On September 3, 2026, OpenAI launched GPT-6 Astra and its president Greg Brockman closed the presentation with “ Welcome to the era of AGI. » The same day, ARC Prize, the organization that manages the ARC-AGI-3 test, measured the model at 99.9% in its best configuration, compared to 30% for the previous best score. In seven weeks, Google responded with Gemini 4 Argon, Anthropic with Claude Fable 5.1 and Opus 5.5, Mistral with a trillion-parameter model, and Qubic published a score obtained without a single graphics card. This comparison ranks seven systems on what can be verified: published scores, price, openness, computing power and governance.
How was this comparison constructed?
Five verifiable criteria, no opinion. The first is the score on ARC-AGI-3a test where the model must learn the rules of unknown games by playing them, designed by François Chollet and administered by ARC Prize on a semi-private set of tasks. We use “verified” scores published on arcprize.organd ARC-AGI-2 when ARC-AGI-3 is missing. The second is the public price of the API per million tokens. The third is openness: downloadable weights or not. The fourth is the declared computing power. The fifth is governance: who decides, who audits.
This framework extends that of our comparison of AI cryptos, which already refused to classify by market capitalization. Here, we refuse to classify by press release.
| System (developer, date) | ARC-AGI-3 | ARC-AGI-2 | API price ($/M tokens, input/output) | Open weights | Calculation and governance |
| GPT-6 Astra (OpenAI, Sept. 3) | 99.9% (provider adapter); 62.7% (standard) | 95.0% | 10/50 | No | More than 100,000 GPUs at Stargate (Texas); first model rated “critical” for cybersecurity; progressive access |
| Gemini 4 Argon (Google, September 30) | 97.3% (provider adapter); 46 to 50% (standard, incomplete executions) | 91.3% | 2/10 at launch, then 4/20 | No | Deployed via the Fairwind program (cyber defenders), then Ultra and API subscribers |
| Claude Fable 5.1 / Opus 5.5 (Anthropic, September 15 and 22) | Unpublished | 90.0% / 93.3% | 10/50 (Fable 5.1) | No | Mythos 5.1, same model with light protection, reserved for approved US organizations |
| Grok 4.7 (xAI, Sep 21) | 10.1% | 61.4% | Not collected | No | Grok 5 announced, no date verified |
| DeepSeek V4.1 Flash (DeepSeek, September 10) | 3.3% | 72.9% | Not collected | Yes (family published in open weights) | Chinese laboratory; best opening/score ratio on ARC-AGI-2 |
| Mistral Large 4 “Le Chonk” (Mistral, Oct. 6) | No public results | No public results | Undisclosed | Promised within three weeks after safety testing | 1 trillion settings, approximately 4,000 Nvidia GPUs; Samsung shareholder, valuation €21 billion |
| Neuraxon / Aigarth (Qubic, September 22) | 0.38 self-declared (0.13 at the beginning), excluding the ARC Prize protocol | Not rated | Not applicable (no LLM API) | Reference open source code | Miner-provided computing (CPU), useful proof of work; scalable architecture, not an LLM |
GPT-6 Astra: the most powerful, the most expensive, the most monitored
The numbers first. On ARC-AGI-3, Astra goes from 62.7% in standard harness to 99.9% when the harness maintains its reasoning state, at an evaluation cost of $18,817 in the best configuration and up to $49,791 in the worst (ARC Prize, September 3). Once the rules of the game are understood, the model uses fewer actions than the human median on 96% of levels. ARC Prize adds two sentences barely mentioned in the press releases: saturate the test “would not represent evidence of achieving AGI,” and the tested environments remain much simpler than the real world.
The model was trained on more than 100,000 GPUs at OpenAI’s Stargate site in Texas, the largest training session in the company’s history, and the first that OpenAI classifies as “critical” for cybersecurity, meaning theoretically capable of finding and exploiting unknown vulnerabilities without step-by-step human guidance (Axios, September 3). Hence progressive access: a restricted circle of organizations, then paying subscribers and the API, patient free users.
Greg Brockman, president of OpenAI, on September 3: “ I think it could be this model. ” Then : ” Welcome to the era of AGI. » Three days later, Nvidia boss Jensen Huang in turn declared that “ the AGI has arrived. » The legal detail matters: the original contract between OpenAI and Microsoft linked an AGI declaration to changes in access rights and revenue sharing; an amendment dated April 27, 2026 removed this clause. Brockman can now speak of AGI as a “ mission concept or spiritual concept » without contractual consequences.
Gemini 4 Argon and Claude: the challengers play another score
Gemini 4 Argonannounced on September 30, is the only other model close to saturation: 97.3% on ARC-AGI-3 in provider adapter, 91.3% on ARC-AGI-2, verified on October 13, with an ARC Prize reserve on incomplete executions in standard harness due to API problems (ARC Prize file). Its weapon: the price: $2 per million incoming tokens and $10 for exit during the launch period, five times less than GPT-6 Astra, then $4 and $20 (9to5Google, September 30).
Google is rolling it out first to cyber defenders via its Fairwind program, with a version lacking cybersecurity protections reserved for trusted defenders.
Anthropic has chosen not to publish an ARC-AGI-3 score for Claude Fable 5.1, launched on September 15, nor for Claude Opus 5.5, launched on September 22. On ARC-AGI-2, the two models are verified at 90.0% and 93.3%, behind the 95% of GPT-6 (ARC Prize file).
Anthropic highlights other measures, long agentic work and scientific tasks in terminal, and charges Fable 5.1 at the same price as Astra, $10 and $50 (Anthropic announcement). A special feature: Claude Mythos 5.1 is the same model with reduced cybersecurity and biology protections, available only to approved US organizations. Three developers, three ways of tackling the same problem, a model too powerful to be open to everyone.
Grok, DeepSeek, Mistral: the race for opening
Behind the first three, the gap is clear. Grok 4.7verified on September 21, reached 10.1% on ARC-AGI-3 and 61.4% on ARC-AGI-2; Grok 5 has been announced for months without a verifiable date. DeepSeek V4.1 Flashverified on September 10, marks 3.3% and 72.9%, but belongs to a family published in open weights, making it the best downloadable model in the table on ARC-AGI-2.
Mistral Large 4 is a different case: no score yet, but a promise. Announced on October 6 under the nickname “The Chonk,” the model has 1 trillion parameters, was trained on approximately 4,000 Nvidia GPUs, “two to three times less than our Chinese competitors” according to Pierre Stock, VP science at Mistral, and its weights must be published within three weeks after security testing (TechCrunch, October 6). Emmanuel Macron sees it as “a third way in AI. » Verification will come the day ARC Prize publishes its file.
The other way: intelligence without a data center
An object of a different nature, compare with caution. Neuraxon is not a great language model. It is a network of simple units that coordinate and specialize, inspired by Marvin Minsky’s “society of the mind”, which the Aigarth evolutionary framework improves from the networks generated by the miners of the Qubic network. The calculation does not come from a data center but from a proof of useful workwhich redirects miners’ power to training, a mechanism detailed in Qubic: Transforming each CPU cycle into real value.
On September 22, Qubic published a score of 0.38 on ARC-AGI-3up from 0.13 at the beginning, after 141 rounds of evolutionary improvements and 450,000 attempts, on ordinary processors, without GPU or integrated language model (recap Qubic). In its analysis of September 26, the project’s scientific team was itself the first to set the limits: the evaluations were carried out on “a few CPUs,” and a distributed network remains “ much less capable on language tasks » than a model trained on 100,000 GPUs (Qubic, Neuraxon Intelligence Academy vol. 13).
Three reading precautions.
The 0.38 is not a score verified by ARC Prize and its protocol does not correspond to that of the semi-private set: it cannot be aligned with Astra’s 99.9%. There is no Neuraxon API comparable to ChatGPT or Claude. And Qubic’s thesis is not ” we are ahead, ” but ” AGI is an emergent, distributed, verifiable organization, not a threshold-crossing model. » This is a position as much on governance as on technique, and it is on this criterion that the project stands out from the table: open reference code, calculation provided by thousands of machines, results presented in conferences rather than in keynotes.
What the ranking doesn’t say
A saturated test stops measuring. ARC-AGI-3 was designed to withstand 2025 models; two systems now exceed it by more than 97%, and ARC Prize warns that this does not prove AGI. Comparing different harnesses is like comparing runners with and without a tailwind: Astra scores 62.7% or 99.9% depending on whether its reasoning state is maintained.
The cost almost never appears in press releases. Evaluating Astra cost up to $50,000 in API calls, while the 500 humans in the benchmark group were paid around $12.78 per game attempted. And OpenAI itself acknowledges that Astra is more difficult to monitor in control evasion tests, a setback the company describes as serious. Amelia Glaese, VP Research at OpenAI: “ When models can do more things autonomously, we should be able to trust them more.“
The next dates are set. Mistral must release Large 4 weights by the end of October; Google expands Gemini 4 Argon to developers; Anthropic has not announced an ARC-AGI-3 transition for Fable 5.1; Qubic will present the state of its useful proof of work during a keynote at the TUM Blockchain Conference in Munich on October 29. The table will be updated with each file verified by ARC Prize.
FAQs
A test where a system must discover for itself the rules of never-before-seen games by playing them, then solve them with as few actions as possible. It measures learning of new skills rather than memorization. It is administered by ARC Prize on a semi-private task set, with a reference group of approximately 500 humans.
On ARC-AGI-3 verified, GPT-6 Astra (99.9%) ahead of Gemini 4 Argon (97.3%). On price, Gemini 4 Argon is five times cheaper at launch. On the opening, DeepSeek today and Mistral Large 4 once its weights are published. No system wins on all criteria.
No. Neuraxon is not a conversational assistant and does not have a consumer API. It is a research program in distributed AI, financed by mining on the Qubic network, the results of which are published by the project and discussed in conferences.
Maximize your Tremplin.io experience with our ‘Read to Earn’ program! For every article you read, earn points and access exclusive rewards. Sign up now and start earning benefits.
