Where it stands
Google has officially announced Gemini 4 Argon, rolling out first to trusted testers in its Fairwind Program, with no launch date for paid API customers or Google AI Ultra subscribers. Later leaks and reports point to a Google Cloud billing listing, a Gemini app model-picker entry and an Antigravity context-window option, but none confirms general availability. Earlier reports about Gemini 4 early post-training and a 'Gemini 4 Pro' comparison leak remain unverified. Next to watch: a public model ID, API documentation and Google's rollout updates.
Leaks point to Argon reasoning-effort settings and an internal Carbon checkpoint
New references to Gemini 4 Argon with low, medium and high reasoning-effort settings have been added recently in Antigravity, according to @testingcatalog. The post says they are hidden in the interface, so they are visible only in the app's code, not to ordinary users.
The same post says the Gemini web app now shows low, medium and high reasoning-effort options for all models, which would unify the reasoning selector across tools. Google has not announced that change.
The post also relays a Business Insider report that Google employees are already testing an internal Gemini 4 checkpoint called Carbon. According to that report, Carbon performs at the Opus 5.5 level on coding tasks. The report is a secondhand account, and Google has not confirmed the testing or the comparison. No benchmark, test method or public model ID has been published.
The post also says that Jetsky, an internal coding tool named in the Business Insider report, is an internal name for Antigravity rather than the IDE version. That would be a possible explanation for the Jetski label in an earlier X listing of placeholder names, but the post does not make that link itself.
For readers, the practical signal is the reasoning-effort setting. It would fit the Thinking level option seen in the Gemini app model picker sighting earlier, but it does not change the access picture: no launch date has been set for paid API customers or Google AI Ultra subscribers. The next confirmations to watch are a public model ID, API documentation and Google's own rollout updates.
Antigravity 2.0 Insiders build shows a Gemini 4 Argon context-window selector
According to @auricxofficial, a code diff of an Antigravity 2.0 Insiders build shows a context-window selector for Gemini 4 Argon, with 256K, 512K and 900K options. Google hasn’t announced these context-window options.
What it shows
The post compares two builds, referred to as main.js and main.now.js. In the newer build, a separate “Context window” list appears only when the base model is Gemini 4 Argon. Other models keep the generic effort menu. The post says the 256K option is the default. It labels 512K as using “about 1.3x quota per turn” and 900K as using “about 1.8x quota per turn,” and both larger options carry an "Expensive" badge.
Image credit: @auricxofficial on X
The post's second image compares the two files. Its table gives the newer build as 147,075 bytes larger (+1.48%), shows no hits for the Gemini 4 Argon string in the older build and one in the newer, and lists a new ChatModelMetadata protobuf message with several field changes. The post notes that raw line diffs are noisy, so it compares extracted identifiers, constants and text instead.
Image credit: @auricxofficial on X
What it could mean
The model itself is not a rumour: Google has announced Gemini 4 Argon, as described in our coverage of the announcement. That announcement says the model is first rolling out to trusted testers through the Fairwind Program, with broader access for paid API customers and Google AI Ultra subscribers afterward, and gives no launch date for either route; its 1 million output-token limit is a separate figure for a single extended run. If the options are real, they would add a per-turn quota cost for long context inside Antigravity, a separate meter from the API token pricing developers pay. Our earlier coverage of Gemini 4 Argon variants reported as Barium-B and a Google Cloud listing belongs to the same leak family.
Status: unconfirmed. We'll update this page when Google confirms or denies it.
Sources
Report says Gemini 4 Argon could launch within a day, with listed prices
An unverified X report said Gemini 4 Argon could launch on the day of the post or the following day, and listed reported prices for global, Flex/Batch and cached input. The report does not confirm that the model launched or that Google made the pricing official.
In a post on X, @pankajkumar_dev says Gemini 4 Argon had a high chance of arriving on the day of the post, or the following day if there was a delay. The supplied source does not provide the post’s publication date, so this timing cannot be converted into a calendar date or treated as a current launch forecast. The post also says the model was already in Google Cloud, presenting that as a reason to expect a near-term launch.
Image credit: @pankajkumar_dev on X
Reported Gemini 4 Argon pricing
The same post lists these reported price lines:
Global: $4 for input and $20 for output.
Flex/Batch: $2 for input and $10 for output.
Global cached input: $0.20/M.
Flex/Batch cached input: $0.10/M.
The source explicitly includes “/M” for the cached-input figures, but it does not define the units for the standard input and output prices. The figures should therefore be treated as reported values rather than a complete or confirmed pricing schedule.
The post does not explain what products or APIs would expose the reported global and Flex/Batch options. It presents “Flex/Batch” as a pricing label without details about eligibility, performance, request handling or availability.
What remains unconfirmed
The supplied report does not confirm that Gemini 4 Argon launched. It provides a prediction tied to the day of the post and the following day, rather than an official release date or availability notice from Google. Its statement that the model was already in Google Cloud is also part of the report, not independent confirmation of public access.
The available information is therefore enough to document the reported timing and prices, but not to establish how users could access the model or how the rates would apply. No official Google announcement, documentation or independent evidence is supplied to confirm the launch or show that the reported prices are final.
The supported conclusion is limited: the report described Gemini 4 Argon as nearing a possible launch and circulated the listed pricing figures, but it did not establish confirmed availability or final pricing.
Sources
Gemini 4 Argon appears in a Vertex AI billing catalogue
Gemini 4 Argon has reportedly appeared in Google Cloud’s Vertex AI billing catalogue, according to a post from @LuminaBench. The listing shows prices for online and Flex/Batch billing, along with supported input and output modalities. It does not, by itself, confirm that Google has publicly launched the model or made it available through an API.
The report is based on a billing-catalogue entry labelled “Gemini 4 Argon Cyber Permissive.” The accompanying image describes the entry as a Google Cloud Vertex AI listing and explicitly says that the release, a public model ID and API access remain unconfirmed.
Image credit: @LuminaBench on X
What the reported Google Cloud listing shows
The image lists two billing categories: global online and Flex/Batch. It also lists text, audio and image input, with text output. Those details describe the catalogue entry, but they do not establish that customers can currently send requests to the model.
The post also says that models appearing in this part of Google Cloud normally launch the same day. That is the poster’s expectation based on prior listings, not a Google announcement or a confirmed launch schedule. The source also notes that something could change before any release.
Reported Gemini 4 Argon pricing
The following figures are the rates displayed in the image. They are reported catalogue figures, not independently verified pricing.
Billing category | Input rate per 1M tokens | Output rate per 1M tokens | Cached input rate per 1M tokens |
|---|---|---|---|
Global online | $4 | $20 | $0.20 |
Flex / Batch | $2 | $10 | $0.10 |
“Cached input” refers to input tokens that are reused through an applicable caching arrangement rather than processed as new input. The image does not specify the eligibility rules, regional availability, quotas, service-level conditions or other billing requirements for either category.
The displayed rates also do not show what a completed request would cost in practice. That would depend on the amount of input and output, whether cached-input pricing applies, and whether a customer can access the relevant billing mode.
Why the listing does not confirm a public launch
A billing-catalogue appearance is a different stage from a public model release. A catalogue can show that a product or billing entry has been prepared without providing customers with a usable endpoint or general access.
The report does not identify an official Google Cloud or Google announcement confirming Gemini 4 Argon. It also does not provide a public model identifier, endpoint, API documentation or access instructions. Without those details, developers cannot treat the catalogue image alone as confirmation that Gemini 4 Argon is available for production use.
The narrower conclusion is that a reported Google Cloud billing entry appears to describe Gemini 4 Argon and gives indicative billing categories and rates. It does not establish when the model will launch, whether the entry is currently usable, or whether access is restricted to a particular group of customers.
What to wait for before treating it as API-accessible
Developers looking for practical access should wait for supporting documentation that answers at least these questions:
What public model ID should an API request use?
Which Vertex AI endpoint or interface supports the model?
Is access available to all customers, or does it require an invitation, subscription or other eligibility condition?
Which regions, quotas and billing modes apply?
Does the listed support for text, audio and image input describe an available API capability or only the catalogue entry?
Until those details are published or otherwise confirmed by an authoritative Google source, the reported listing is best understood as a pricing and availability signal—not proof that Gemini 4 Argon has publicly launched or that its API is open.
Sources
Gemini 4 Argon reportedly appears in the Gemini app model picker
Gemini 4 Argon has reportedly appeared in the Gemini app’s model picker for at least one user, but the report is unverified and does not establish that Google has launched the model or begun a public rollout.
In a post on X, @thtbee_ says a user on a Chinese site saw the entry in the Gemini iOS app. The post also says the report could not be verified, leaving open the possibility that the screenshot is fake or that the entry does not represent a public release.
What the report says about Gemini 4 Argon
The reported sighting concerns a model-picker option labeled “4 Argon.” The post describes it as appearing alongside other Gemini model options and suggests that, if the report is genuine, a rollout could be only a few days away. That timing is speculation from the post, not a release date or schedule announced by Google.
The available material does not show whether the model-picker entry appeared for more than one user. It also does not establish whether “4 Argon” is a public model, a temporary test, an internal label, or an altered image.
What the attached model-picker screenshot shows
The supplied screenshot appears to show the Gemini app with a model picker containing “3.5 Flash-Lite,” “3.8 Flash,” “3.1 Pro,” and a checked “4 Argon” option. The “4 Argon” entry has the description “Complex tasks” beneath it. The menu also includes a “Thinking level” option.
Those labels are the visible evidence in the image, but they do not independently confirm the model’s identity, capabilities, or availability. In particular, “Complex tasks” is a menu description and should not be treated as a measured performance result.
Image credit: @thtbee_ on X
What remains unconfirmed about availability
The supplied post and screenshot do not confirm that Gemini 4 Argon is available to the public, limited to selected users, or scheduled for release. No official Google announcement, documentation, release date, regional availability information, or independent confirmation was supplied with the report.
For now, the most supportable description is an unconfirmed Gemini app sighting. Readers should not treat the screenshot as proof that Gemini 4 Argon has launched or that a rollout is imminent.
Sources
Four Barium-B and Jetski placeholder labels surface in an X listing
The labels have been described as possible leaked Gemini 4 Argon placeholders, but @lyraxana’s X post itself only lists four entries and does not name Gemini 4 Argon. The listing includes one Jetski-labelled slot and three Barium-B variants whose names mention a tool-leakage fix, summaries and two medium-effort values.
The post does not confirm that these entries are public models, official Gemini 4 Argon components or tested configurations. It provides no benchmarks, access instructions, pricing, launch timing or official confirmation.
What the reported placeholder list contains
The four entries are:
lsp-slot1-lc-jetski— labelled “LSP Teamfood Slot 1”barium-b-tool-leakage-fix-tf— described as a Barium-B model with a tool-leakage fixbarium-b-med-is-p20-tf— described as a Barium-B model with summaries and medium effort at 0.20barium-b-med-is-p50-tf— described as a Barium-B model with summaries and medium effort at 0.50
The first entry does not include “Barium-B” in its listed name. The post does not explain what “LSP,” “Teamfood” or “Jetski” means, so those terms should remain labels rather than being presented as confirmed product names or capabilities.
The reported Barium-B changes
One Barium-B entry is identified by a reported tool-leakage fix. The label names a reported tool-leakage fix, but it does not establish how the fix works or whether it improves performance. The listing provides no test results or explanation of the underlying issue.
The other two Barium-B entries are described as having summaries and a medium-effort setting. The post does not define what the summaries contain or explain what the effort setting controls.
How the two medium-effort labels differ
The summary-enabled entries appear to represent two separate reported configurations:
barium-b-med-is-p20-tfuses the listed medium-effort value 0.20.barium-b-med-is-p50-tfuses the listed medium-effort value 0.50.
The values distinguish the configurations in the placeholder list, but they do not show that one version is more capable, accurate or reliable than the other. The supplied post includes no benchmarks, technical notes or test results that would support a performance ranking.
What the listing does not confirm
The supplied X listing does not establish:
whether these labels are associated with Gemini 4 Argon;
whether Gemini 4 Argon has been officially announced;
whether Barium-B is an official model family;
whether the entries are leaked internal placeholders, experimental configurations or publicly accessible models;
what “LSP Teamfood” means;
what the Jetski label describes;
what the tool-leakage fix changes in practice;
how the summary feature works; or
what the medium-effort values control.
The safest reading is that the post lists model-like labels and named configuration differences. The possible connection to leaked Gemini 4 Argon placeholders comes from the supplied editorial framing, not from the text of the post itself. The evidence does not confirm a release, access details, capabilities or performance.
Source
Google says Argon agents handle internal data-center and migration work
Google says its Gemini 4 Argon agents are being used for specialized engineering work inside the company, including data-center memory optimization, large-scale C and C++ to Rust migrations, and performance tuning. The reported examples concern Google’s internal infrastructure and workflows—not a public product launch or evidence that the agents operate without human oversight.
In a post on X, @kimmonismus reported that Google’s agents had analyzed infrastructure profiling data, identified memory optimizations and helped apply changes that were later rolled out. The post also described a one-million-token output limit, but did not specify whether that limit applies to every Argon agent or identify the relevant model configuration.
The supplied image presents these activities as part of broader internal work involving specialized coding, research and engineering productivity.
Image credit: @kimmonismus on X
Optimizing memory across Google’s data centers
Profiling telemetry is operational data that records how systems use resources. In this case, the reported workflow involved analyzing fleet-wide profiling data to find opportunities to reduce memory use, then applying changes across Google’s data centers.
The X report says those changes freed more than 300 tebibytes (TiB) of memory after rollout. A tebibyte is a unit of digital storage equal to 1,024 gibibytes. The supplied image gives a larger estimate of 500 TiB to 1 pebibyte (PiB) in total savings, but does not provide a measurement method, timeframe or baseline for that estimate.
Those figures should therefore be read as Google’s reported results rather than independently verified measurements. The material also does not explain which systems were changed, how agents proposed the optimizations, or what human review took place before deployment.
Rewriting C and C++ code in Rust
The reported work also includes migrations from C and C++ to Rust. A migration of this kind rewrites existing software in another programming language while attempting to preserve its behavior. Rust is often used for systems programming because its compiler can enforce important memory-safety rules, but moving a large, mature codebase remains a substantial engineering task.
The supplied X post describes more than 800,000 lines of kernel code, while the accompanying image says the migrations range from tens of thousands of lines in core libraries such as re2 and libgav1 to that larger kernel project. The image identifies the kernel project as the Fuchsia OS Zircon kernel and says the work includes automated and manual auditing, emulation testing and review before changes reach production.
That description indicates a staged engineering process rather than a simple instruction to rewrite an entire codebase and deploy the result. It also does not establish that the full 800,000-plus-line migration is complete. The source gives no further details about which portions have been migrated, how compatibility was measured or how much of the work was generated by agents versus completed by engineers.
Tuning performance-critical code
One example in the report concerns a video decoder. Google reportedly used agents to replace 32,000 lines of SIMD code with safe Rust. SIMD, or single instruction, multiple data, is a technique for applying one operation to many data elements at once and is commonly used in performance-sensitive software.
According to the supplied post, the existing Rust port became 2.7 times faster while producing identical video output. The post does not specify the test hardware, workload, baseline measurements or evaluation procedure, so the result cannot be treated as a general benchmark for Argon or for Rust-based video decoders. It is a reported result from one internal example.
Other reported engineering uses
The supplied image also describes Argon helping Google’s quantum-computing researchers optimize the “spacetime” resources used by quantum subroutines. In this context, the phrase refers to the combined resource cost represented by the number of qubits and the number of operations, or gates, used by an algorithm.
The image says one example beat a published baseline by 40% in minutes. It does not identify the baseline, workload, evaluation setup or publication, so the reported comparison lacks the detail needed to judge how broadly the result applies.
What the report establishes—and what it does not
Taken together, the examples portray Argon as being used for bounded, high-value engineering tasks: analyzing operational data, proposing or implementing code changes, optimizing performance and supporting large migrations. The reported safeguards include automated and manual auditing, emulation testing and review before production rollout.
The material does not establish that Gemini 4 Argon is generally available, that it can independently manage Google’s infrastructure, or that every reported change was completed without engineers. It also does not provide primary technical documentation, dates, baselines or detailed outcome measurements.
The most supportable conclusion is narrower: Google is reported to be applying its agents to difficult internal engineering workflows, where the systems can work over large codebases and infrastructure data but where testing, auditing and deployment review remain important parts of the process.
Sources
Artificial Analysis chart places Argon at 53, level with GPT-6 Astra
An X post from @LuminaBench reports that Gemini 4 Argon received a score of 53 on the Artificial Analysis Intelligence Index. The supplied chart shows Gemini 4 Argon (high) and GPT-6 Astra (max) with the same displayed score, while a separate cost-versus-index plot places Gemini 4 Argon near the chart’s dotted Pareto line.
The post and image do not independently confirm Gemini 4 Argon’s availability, the benchmark methodology, its pricing, or how it performs in ordinary use.
What the post reports about Gemini 4 Argon
The post says Gemini 4 Argon scored 53 on the Artificial Analysis Intelligence Index and was level with GPT-6 Astra. In the image, both models have bars labelled 53 in the upper comparison chart. The post also describes Gemini 4 Argon as being “basically right on the Pareto frontier” for cost.
“Level” here refers to the displayed index score, not proof that the two models produce equivalent results on every evaluation or task. The post does not include the underlying scores, test conditions, or results needed to make that broader comparison.
The thread also includes the poster’s personal statements that they tested the model and found it strong and good to use. Those comments do not include a test setup, reproducible examples, or measured results, so they are not enough to independently assess the model’s practical performance.
What the supplied chart displays
The upper portion of the attached Artificial Analysis chart labels Gemini 4 Argon (high) with a score of 53. GPT-6 Astra (max) is also labelled 53. The same chart shows other models at different displayed index values, but the report does not provide the data behind those scores.
The chart also includes a “Not publicly available” annotation. That is a label displayed in the supplied image, not independent confirmation of Gemini 4 Argon’s release or distribution status.
The lower portion is a separate chart titled “Intelligence Index vs. Cost per Intelligence Index Task.” Its vertical axis represents the Intelligence Index, while its horizontal axis shows cost per task in US dollars on a logarithmic scale. Gemini 4 Argon (high) appears at an index value around 53 and close to the dotted Pareto line.
Image credit: @LuminaBench on X
The chart therefore presents two related but different views: the upper bars compare displayed index scores, while the lower plot relates index values to estimated cost per task. Seeing Gemini 4 Argon near the dotted line does not by itself establish an exact price, a best-in-class cost position, or a recommendation for production use.
How to read the reported cost-performance position
A Pareto line is a comparison frontier: points near it can represent more attractive trade-offs between the two plotted measures than points that are clearly worse on both. In this image, the relevant measures are the displayed Intelligence Index and cost per Intelligence Index task.
The image places Gemini 4 Argon near that dotted line, which supports the narrower observation that it appears close to the chart’s illustrated cost-performance frontier. It does not show enough information to determine the exact cost for Gemini 4 Argon or to reproduce the comparison.
The chart labels the measure as a weighted average cost in U.S. dollars per Artificial Analysis Intelligence Index task, but it does not show the underlying pricing assumptions or explain whether that estimate is based on API prices or another cost calculation.
Evaluations named in the chart
Text at the top of the image says that Artificial Analysis Intelligence Index v4.3.2 incorporates 10 evaluations. The image lists:
AA-Briefcase v1.1
GDPval-AA v2.1
AutomationBench-AA
Terminal-Bench 4.0
SciCode
Humanity’s Last Exam
GDP.pdf
CritPt
AA-Omniscience
AA-LCR v1.1
The image does not show the weighting, scoring rules, individual results, or definitions for these evaluations. As a result, the overall score of 53 cannot be unpacked into conclusions about any one capability from the image alone.
What this report does not confirm
The post and chart do not establish that Gemini 4 Argon is officially released or publicly available. No official Google announcement, product page, release information, or API documentation is supplied here.
They also do not independently verify that Gemini 4 Argon matches GPT-6 Astra beyond the two models sharing a displayed score of 53. A shared aggregate score can conceal different results across the underlying evaluations.
Finally, the material does not confirm a precise cost ranking or real-world superiority. The most supportable reading is narrower: the X post reports a score of 53, the attached chart displays Gemini 4 Argon and GPT-6 Astra at the same index value, and the lower chart places Gemini 4 Argon near its illustrated Pareto line. Further conclusions require the methodology, underlying results, pricing assumptions, and an official or reproducible account of the model.
Sources
Google announces Gemini 4 Argon for trusted testers first
Google has announced Gemini 4 Argon, a frontier model designed for complex, long-running workflows in software engineering, enterprise knowledge work and cybersecurity defense. In a post on X, @GoogleDeepMind says the model is initially rolling out to trusted testers through its Fairwind Program rather than launching broadly.
Image credit: @GoogleDeepMind on X
Google says broader access will begin with paid API customers and Google AI Ultra subscribers after the early testing phase. It has not announced a launch date or detailed eligibility requirements for either route.
What Gemini 4 Argon is designed to do
Google positions Argon around three main areas: software engineering, enterprise knowledge work and cybersecurity defense. The company also describes capabilities in reasoning, multimodal understanding and creative writing, but its examples focus most heavily on coding and security.
The announcement says thousands of Google employees are already using Argon for specialized coding, research and writing tasks. Google presents these as internal examples of the model’s potential, not as typical results for external users.
What the 1 million-token output limit changes
A token is a unit of text or other model input and output; the amount represented by a token varies by language and content. Google says Gemini 4 Argon can produce up to 1 million output tokens in a single trajectory, up from a previous limit of 64,000 tokens.
That limit is relevant to workflows in which the model must sustain a long sequence of analysis and actions rather than produce a short answer. In principle, a larger output allowance can give an agent more room to inspect a codebase, run or interpret experiments, revise an approach and document the result in one extended process. It does not by itself show that the model will solve every long task reliably.
Image credit: @GoogleDeepMind on X
Google’s description of the capability is tied to “long-horizon” work: problems that involve multiple steps and require the model to maintain a coherent plan over time. The practical value will depend on factors such as tool access, safeguards, context quality and how well the model’s intermediate work can be checked.
Coding and large codebase migrations
Google says Argon agents are being used for everyday debugging, algorithm design and large migrations from C and C++ to Rust. The company says those migrations range from tens of thousands of lines in libraries such as re2 and libgav1 to more than 800,000 lines in the Fuchsia OS Zircon kernel.
The announcement gives a more specific example involving libgav1, Google’s open-source video decoder. Google says Argon agents replaced 32,000 lines of SIMD code in an existing Rust port through repeated profile-guided experiments and compiler analysis. The resulting decoder reportedly ran 2.7 times faster than that Rust port while producing identical video output.
That is an internal Google example, not a promise about the result a typical developer will obtain. Google also says large migrations undergo automated and manual auditing, emulation testing and review before production deployment, underscoring that the model’s output still requires engineering controls.
The company additionally says Argon helped quantum-computing researchers optimize a subroutine’s spacetime resources—qubits multiplied by gates—and beat a published baseline by 40% in one example. Google reports that Argon agents also analyzed fleet-wide profiling data and identified memory optimizations that freed more than 300 tebibytes of memory once deployed, with an estimated 500 tebibytes to 1 pebibyte in total savings.
Enterprise knowledge work and multimodal tasks
Google says Argon is intended for knowledge work in areas including finance, legal research and drafting, tax and business automation. It reports leading results on the Vals Index, which measures work across finance, coding, legal and tax tasks, as well as on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark.
The company also reports a first-place result on AutomationBench, a Zapier benchmark for end-to-end execution across core business functions, with a score of 51.3%.
For work involving visual material, Google says Argon can analyze professional charts, identify details in long videos and act on a sequence of documents. It reports a 91.7% score on LVBench, an evaluation of long-video understanding, and a 71.6% score on Chartography.
These figures are Google-reported evaluation results. They do not establish how Argon would perform on a particular company’s documents, software or business processes.
Cybersecurity defense
Google says it trained Argon to find, validate and patch software vulnerabilities. The company plans to release the model without cyber guardrails to trusted defenders and its own internal teams so they can use its full cybersecurity capabilities.
Google also says Wiz is using Argon through its Scan for Good initiative, which aims to find and remediate high-risk exposures in critical public infrastructure. In one early demonstration, Google reports that the model found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide. The announcement does not provide detailed technical information about the vulnerability or the remediation process.
On CWE-bench v1, which evaluates vulnerability remediation, Google reports that Argon tied for first place with a score of 68%. The company also describes improvements over its earlier 3.8 Flash Cyber model on internal vulnerability-discovery and black-box penetration-testing evaluations.
Because these capabilities can be dual-use, the announcement’s cybersecurity access restrictions are part of the product story. Argon’s initial rollout is aimed at trusted testers and cyber defenders, not unrestricted public use.
How to read the reported benchmarks
Google’s benchmark comparison shows Argon alongside GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across selected evaluations. The table below reproduces the figures shown in that comparison.
Evaluation | Area | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|---|
Vals Index | Knowledge work | 68.9% | 63.1% | 65.8% | 67.0% |
AutomationBench | Knowledge work | 51.3% | 41.4% | 31.4% | 42.5% |
DeepSWE v1.1 | Agentic coding | 77.9% | 74.1% | 67.4% | 74.2% |
Chartography | Multimodal understanding | 71.6% | 71.0% | 46.2% | 66.3% |
LVBench | Multimodal understanding | 91.7% | 87.5% | 79.7% | 83.7% |
CWE-bench v1 | Cybersecurity | 68.0% | 68.0% | 58.0% | 67.0% |
Google DeepMind reports these figures, and they have not been independently verified here. The announcement does not provide the complete prompts, model configurations, test conditions or validation process needed to determine how directly the scores can be compared across every evaluation. A higher score on one benchmark also does not establish broader superiority across all enterprise, coding or security tasks.
The reported results are most useful as indicators of the tasks Google is prioritizing: multi-step coding, business automation, visual understanding and vulnerability remediation. They are not substitutes for testing a model against an organization’s own data, tools and review requirements.
Pricing and access
Google says Gemini 4 Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced at 95% below the input-token price, according to the announcement.
The announcement references a pricing footnote but does not spell out additional conditions, limits or product-specific details. It also does not say when the introductory prices will change.
Access is currently staged. Google says Argon is rolling out to a set of trusted testers through the Fairwind Program, with the initial announcement emphasizing trusted cyber defenders. Google has not published the program’s eligibility rules or application process in the supplied announcement.
Google says it plans to expand access to developers, enterprises and consumers after gathering feedback and strengthening its safeguards. The company says that broader release will start with paid API customers and Google AI Ultra subscribers, but no specific date or detailed access requirements are provided.
Safeguards before broader release
Google says it is continuing work on four areas of frontier safeguards before making Argon more widely available:
Misuse prevention: The model is designed to refuse harmful cyber and chemical, biological, radiological and nuclear requests while supporting legitimate dual-use research. Google says internal and external red teams tested these safeguards.
Prompt-injection resistance: Google describes protections against indirect prompt injection, in which malicious instructions hidden in documents or other context attempt to redirect the model.
Misalignment monitoring: Google says it monitors the model’s reasoning and actions and can stop execution when behavior moves beyond the user’s intent.
Hardened environments: The company says it is isolating and sealing sandboxed environments before high-risk training or evaluations and plans to share related agent-security practices with partners.
These measures help explain why Google is describing Argon as a phased announcement rather than a general launch. The model’s reported capabilities are aimed at workflows where mistakes, unauthorized actions or successful misuse could have significant consequences. For now, Gemini 4 Argon is best understood as a limited-access system undergoing evaluation, with broader availability and real-world performance still to be established.
Sources
Gemini 4 Pro comparison leak claims higher benchmarks and lower token prices
In a post on X, @JustLingonberry presents an alleged Gemini 4 Pro leak as a major competitive threat to Anthropic and OpenAI models. The accompanying comparison image shows Gemini 4 Pro with higher displayed figures than Claude Opus 5.5 and GPT-6 Astra on each listed benchmark, along with lower displayed token prices. However, the material does not include an official Google announcement, model documentation, testing methodology or pricing page.
The thread is therefore best read as a report of what one comparison image appears to show—not as confirmation that Gemini 4 Pro has launched, achieved the listed results or will be sold at those prices.
Image credit: @JustLingonberry on X
What the X thread is reporting
The opening post says that Gemini 4 Pro “leaks” dominate both Anthropic and OpenAI and suggests that competing models may need to match its pricing. Replies add the author’s view that the leak appears believable and that the model may match or exceed Claude Opus 5.5.
Those statements are interpretations from the thread, not results from a disclosed evaluation. The thread also speculates that Google’s control over its own Tensor Processing Units (TPUs)—specialized chips for running AI workloads—could help with speed and pricing, but it supplies no evidence showing that the chips produced an advantage. There are also no measurements of inference speed, operating costs or margins.
The opening post refers to “Fable 5.5” and “GPT 6.1,” while the comparison image labels the competitors Claude Opus 5.5 and GPT-6 Astra. The source does not establish whether these are alternate names for the same models or different references.
Reported benchmark comparison
The attached image lists the following figures:
Benchmark | Gemini 4 Pro (leak) | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
DeepSWE v1.1 | 88.7 | 74.2 | 74.1 |
Terminal-Bench 2.1 | 95.3 | ~88 | ~88 |
OSWorld 2.0 | 86.8 | 81.8 (partial) | 72.6 (offline partial) |
GDPval-AA | 2,064 (v2) | 1,846 (v2.1) | 1,542 (v2.1) |
The image displays a higher Gemini figure in every listed row. That describes the comparison graphic, not a verified model lead: the rows use different benchmark versions and include qualifications such as “partial” and “offline partial,” while the source provides no hardware, prompts, software versions, test dates, sampling settings, scoring process or evidence that the results were reproduced. Those differences and missing details mean the figures cannot be treated as a verified or necessarily comparable cross-model advantage.
The comparison image is evidence of what was displayed in the leak, not independent evidence that the figures are accurate.
Displayed context-window and token-price figures
A context window is the token limit for the context a model can process in a request or conversation. Depending on the system, that limit may include some or all of the generated output as well as the conversation or other input context. The image labels Gemini 4 Pro’s context window as 2 million tokens, compared with 1 million tokens for Claude Opus 5.5 and 1.1 million tokens for GPT-6 Astra. The supplied thread later says that Gemini 4 Pro is “confirmed to be 1M context not 2M,” so the alleged Gemini figure remains unresolved.
The image also displays these prices per 1 million tokens:
Gemini 4 Pro: $2.25 / $11.25
Claude Opus 5.5: $4 / $20
GPT-6 Astra: $10 / $50
The image does not explain what the two prices in each pair represent. It does not identify input and output billing, service tiers, caching, batch discounts, regional terms or whether the figures are proposed, unofficial or final. Readers should not treat the displayed amounts as confirmed API pricing.
Neither the image’s 2-million-token label nor the thread’s 1-million-token correction is independently verified by the supplied source.
What the comparison does—and does not—show
If the displayed figures are accurate and comparable, they would suggest a potentially strong benchmark and price position for the alleged Gemini model. They do not show how the models perform on a particular user’s workload, how quickly they generate responses, or what their effective cost would be in production.
A benchmark score can depend on the task definition, evaluation harness, prompts and scoring rules. A displayed token price also cannot be compared reliably without knowing which type of tokens each number covers and whether the services use equivalent conditions. The thread’s predictions about competitive pricing, margins, speed and market effects remain speculation.
Open questions about the leak
The leak leaves several material questions unanswered:
Has Google officially announced Gemini 4 Pro or published documentation for it?
Who produced the comparison and under what test conditions?
Were the benchmark results independently reproduced?
Are the model names in the image consistent with the names used in the thread?
Is the context window 2 million tokens, 1 million tokens or something else?
Do the displayed prices represent official, final API rates, and what do the slash-separated figures mean?
Is the alleged model available to users, and if so, through which service?
Until those questions have answers from reliable documentation or reproducible testing, the comparison is useful mainly as an account of what the leak is reporting. It should not be treated as confirmation that Gemini 4 Pro has launched, definitively beats Claude or OpenAI models, or will force competitors to change their prices.
Sources
Gemini 4 early post-training version reportedly planned as soon as possible
In a post on X, @LuminaBench reports that Google DeepMind chief Koray Kavukcuoglu said the team intends to roll out an early post-training version of Gemini 4 as soon as possible. Text shown with the report says Kavukcuoglu hopes Google can ship the flagship model “much earlier” than the end of the year.
That points to an intention to move quickly, not to a confirmed Gemini 4 release date. The available evidence does not establish when an early version or the final model will become publicly available.
What is reported about Gemini 4’s early version
The supplied report text describes Gemini 4 as having entered an early phase of post-training. In machine learning, post-training is the stage after a model’s initial training in which it can be further adjusted, such as to improve instruction following or its responses to users. The report does not establish which post-training methods Google is using or what capabilities this version has.
The attached text quotes Kavukcuoglu as saying: “Our intention is to roll out an early post-training version as soon as possible, because we’ve already seen promising results and are very excited.” The same text says he hopes to ship the flagship model “much earlier” than the end of the year.
Those remarks describe a plan and an expectation. They do not confirm that Google has completed the work, publicly released the tested checkpoint, opened access to users or fixed a public launch schedule.
Is there a confirmed Gemini 4 release date?
No confirmed date is established by the available evidence. “As soon as possible” describes urgency rather than a calendar date, while “much earlier” than the end of the year provides only a broad expectation. It does not specify a month, day or public availability window.
The supplied report is attributed to The Information’s AI Agenda Live Summit coverage, but it does not cite a firm public launch schedule or access terms. The wording also leaves open whether the early post-training version would be a limited checkpoint, an evaluation release or a broadly available product.
Nothing in the report establishes its access terms, pricing, capabilities or relationship to the final Gemini 4 model.
What the arena assessment does—and does not—show
The X post says its author tested a checkpoint in an arena and personally considered the model highly capable. That is the author’s assessment, not independently measured benchmark evidence. The post does not provide arena scores, the test setup, a comparison set or reproducible results.
The available evidence also does not establish how LuminaBench accessed the checkpoint, whether it came directly from Google or whether it was publicly available. The assessment may explain why the author is optimistic, but it does not establish Gemini 4’s performance or confirm that the checkpoint represents the final model.
What remains unknown
The available report does not answer several practical questions about the Gemini 4 early post-training release:
When, if ever, the early version will become available.
Whether it will be accessible to the public, selected testers or only internal evaluators.
What capabilities or limitations distinguish it from the final model.
Whether Google will charge for access or offer it through an existing product or service.
When the flagship Gemini 4 model will launch.
For now, the evidence supports a narrower conclusion: Google DeepMind has reportedly expressed an intention to roll out an early post-training version as soon as possible, and its chief reportedly hopes to ship the flagship model before the end of the year. Neither statement establishes a firm Gemini 4 release date or confirms that an early version is already publicly available.
0 comments
No approved comments yet. You can start the conversation.
Leave a comment