Gemini

<p>On 3 September 2026, OpenAI launched GPT-6 Astra and its president Greg Brockman closed the presentation with “<em>Welcome to the AGI era.</em>” The same day, ARC Prize, the organization that runs the ARC-AGI-3 test, measured the model at 99.9% in its best configuration, against 30% for the previous top score. Within seven weeks, Google answered with Gemini 4 Argon, Anthropic with Claude Fable 5.1 and Opus 5.5, Mistral with a 1-trillion-parameter model, and Qubic published a score obtained without a single graphics card. This comparison ranks seven systems on what can be checked: published scores, prices, openness, compute and governance.</p>
<div class="in-brief">
 <div class="in-brief--container"><span></span><span></span>
  <h2 class="in-brief--title" id="in-brief">In Brief</h2>
  <ul>
   <li>ARC-AGI-3, the reference test for general intelligence, is close to saturated: GPT-6 Astra at 99.9% and Gemini 4 Argon at 97.3% with the “provider adapter” harness, both verified by ARC Prize.</li>
   <li>Anthropic has no published ARC-AGI-3 score for Claude Fable 5.1 or Opus 5.5; on ARC-AGI-2, Opus 5.5 reaches 93.3% and GPT-6 95%.</li>
   <li>Grok 4.7 (10%) and DeepSeek V4.1 Flash (3.3%) remain far behind on ARC-AGI-3; Mistral Large 4, announced on 6 October with 1 trillion parameters, has no public result yet.</li>
   <li>Qubic claims 0.38 on ARC-AGI-3 for its Neuraxon model, up from 0.13 at the start, on ordinary processors: a self-reported figure, outside the ARC Prize protocol.</li>
  </ul>
 </div>
</div>
<h2 id="h-how-was-this-comparison-built">How was this comparison built?</h2>
<p><strong>Five verifiable criteria, no opinions.</strong> The first is the score on <strong>ARC-AGI-3</strong>, a test where the model has to learn the rules of unknown games by playing them, designed by François Chollet and administered by ARC Prize on a semi-private task set. We use the “verified” scores published on <a href="https://arcprize.org/results" rel="noopener noreferrer nofollow">arcprize.org</a>, and ARC-AGI-2 where ARC-AGI-3 is missing. The second is the public API price per million tokens. The third is openness: downloadable weights or not. The fourth is declared training compute. The fifth is governance: who decides, who audits.</p>
<div class="my-4">
 <div class="ads "></div>
</div>
<p>This framework extends the one in our <a href="https://www.cointribune.com/en/top-ai-crypto-in-2026-why-comparing-qubic-tao-render-fet-and-near-by-price-no-longer-makes-sense/" rel="noopener noreferrer nofollow">comparison of AI cryptos</a>, which already refused to rank by market cap. Here, we refuse to rank by press release.</p>
<figure class="wp-block-table">
 <table>
  <tbody>
   <tr>
    <td><strong>System (developer, date)</strong></td>
    <td><strong>ARC-AGI-3</strong></td>
    <td><strong>ARC-AGI-2</strong></td>
    <td><strong>API price ($/M tokens, input / output)</strong></td>
    <td><strong>Open weights</strong></td>
    <td><strong>Compute and governance</strong></td>
   </tr>
   <tr>
    <td>GPT-6 Astra (OpenAI, 3 Sept.)</td>
    <td>99.9% (provider adapter); 62.7% (standard)</td>
    <td>95.0%</td>
    <td>10 / 50</td>
    <td>No</td>
    <td>Over 100,000 GPUs at Stargate (Texas); first model rated “critical” for cybersecurity; staged access</td>
   </tr>
   <tr>
    <td>Gemini 4 Argon (Google, 30 Sept.)</td>
    <td>97.3% (provider adapter); 46 to 50% (standard, incomplete runs)</td>
    <td>91.3%</td>
    <td>2 / 10 at launch, then 4 / 20</td>
    <td>No</td>
    <td>Rolled out through the Fairwind program (cyber defenders), then Ultra subscribers and API</td>
   </tr>
   <tr>
    <td>Claude Fable 5.1 / Opus 5.5 (Anthropic, 15 and 22 Sept.)</td>
    <td>Not published</td>
    <td>90.0% / 93.3%</td>
    <td>10 / 50 (Fable 5.1)</td>
    <td>No</td>
    <td>Mythos 5.1, same model with lighter safeguards, reserved for vetted US organizations</td>
   </tr>
   <tr>
    <td>Grok 4.7 (xAI, 21 Sept.)</td>
    <td>10.1%</td>
    <td>61.4%</td>
    <td>Not collected</td>
    <td>No</td>
    <td>Grok 5 announced, no verified date</td>
   </tr>
   <tr>
    <td>DeepSeek V4.1 Flash (DeepSeek, 10 Sept.)</td>
    <td>3.3%</td>
    <td>72.9%</td>
    <td>Not collected</td>
    <td>Yes (family released as open weights)</td>
    <td>Chinese lab; best openness-to-score ratio on ARC-AGI-2</td>
   </tr>
   <tr>
    <td>Mistral Large 4 “Le Chonk” (Mistral, 6 Oct.)</td>
    <td>No public result</td>
    <td>No public result</td>
    <td>Not disclosed</td>
    <td>Promised within three weeks after safety testing</td>
    <td>1 trillion parameters, around 4,000 Nvidia GPUs; Samsung as shareholder, €21bn valuation</td>
   </tr>
   <tr>
    <td>Neuraxon / Aigarth (Qubic, 22 Sept.)</td>
    <td>0.38 self-reported (0.13 at start), outside the ARC Prize protocol</td>
    <td>Not evaluated</td>
    <td>Not applicable (no LLM API)</td>
    <td>Open reference code</td>
    <td>Compute supplied by miners (CPUs), useful proof of work; evolutionary architecture, not an LLM</td>
   </tr>
  </tbody>
 </table>
 <figcaption><em>Sources: ARC Prize (verified results pages, accessed 21 October 2026), Axios, 9to5Google, Anthropic, TechCrunch, Qubic. “Provider adapter” ARC-AGI-3 scores keep the reasoning state between actions; the “standard” harness starts from scratch at every action.</em>
 </figcaption>
</figure>
<h2 id="h-gpt-6-astra-the-strongest-the-priciest-the-most-closely-watched">GPT-6 Astra: the strongest, the priciest, the most closely watched</h2>
<p>The numbers first. On ARC-AGI-3, Astra goes from 62.7% in the standard harness to <strong>99.9%</strong> when the harness retains its reasoning state, at an evaluation cost of $18,817 in the best configuration and up to $49,791 in the worst (<a href="https://arcprize.org/blog/astra" rel="noopener noreferrer nofollow">ARC Prize, 3 September</a>). Once a game’s rules were understood, the model used fewer actions than the human median on 96% of levels. ARC Prize adds two sentences the press releases barely picked up: saturating the test <em>“would not represent proof of achieving AGI,”</em> and the environments tested remain far simpler than the real world.</p>
<p>The model was trained on more than <strong>100,000 GPUs</strong> at OpenAI’s Stargate site in Texas, the largest training run in the company’s history, and it is the first one OpenAI rates at the “critical” level for cybersecurity, meaning capable in theory of finding and exploiting unknown vulnerabilities without step-by-step human guidance (<a href="https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman" rel="noopener noreferrer nofollow">Axios, 3 September</a>). Hence staged access: a restricted circle of organizations, then paying subscribers and the API, with free users waiting.</p>
<p>Greg Brockman, OpenAI’s president, on 3 September: “<em>I think it might be about this model.</em>” Then: “<em>Welcome to the AGI era.</em>” Three days later, Nvidia chief Jensen Huang declared in turn that “<em>AGI has arrived.”</em> The legal detail matters: the original contract between OpenAI and Microsoft tied an AGI declaration to changes in access rights and revenue sharing; an amendment dated 27 April 2026 removed that clause. Brockman can now speak of AGI as a “<em>mission concept or spiritual concept</em>” with no contractual consequence.</p>
<h2 id="h-gemini-4-argon-and-claude-the-challengers-play-a-different-tune">Gemini 4 Argon and Claude: the challengers play a different tune</h2>
<p><strong>Gemini 4 Argon</strong>, announced on 30 September, 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 13 October, with a caveat from ARC Prize about incomplete runs in the standard harness due to API issues (<a href="https://arcprize.org/results/google-gemini-4-argon" rel="noopener noreferrer nofollow">ARC Prize card</a>). Its weapon is price: $2 per million input tokens and $10 for output during the launch period, five times less than GPT-6 Astra, then $4 and $20 (<a href="https://9to5google.com/2026/09/30/gemini-4-argon-announcement/" rel="noopener noreferrer nofollow">9to5Google, 30 September</a>).</p>
<p>Google is rolling it out first to cyber defenders through its Fairwind program, with a version stripped of cyber safeguards reserved for trusted defenders.</p>
<p><strong>Anthropic</strong> chose not to publish an ARC-AGI-3 score for Claude Fable 5.1, launched on 15 September, or for Claude Opus 5.5, launched on 22 September. On ARC-AGI-2, the two models are verified at 90.0% and 93.3%, behind GPT-6’s 95% (<a href="https://arcprize.org/results/anthropic-claude-fable-5-1" rel="noopener noreferrer nofollow">ARC Prize card</a>).</p>
<p>Anthropic puts forward other measures, long agentic work and scientific tasks in the terminal, and prices Fable 5.1 at the same rate as Astra, $10 and $50 (<a href="https://www.anthropic.com/claude-fable-and-mythos-5-1" rel="noopener noreferrer nofollow">Anthropic announcement</a>). One particularity: Claude Mythos 5.1 is the same model with lighter safeguards in cybersecurity and biology, available only to vetted US organizations. Three developers, three ways of handling the same problem, a model too capable to be opened to everyone.</p>
<div class="my-4">
 <div class="ads "></div>
</div>
<h2 id="h-grok-deepseek-mistral-the-race-to-openness">Grok, DeepSeek, Mistral: the race to openness</h2>
<p>Behind the top three, the gap is clear. <strong>Grok 4.7</strong>, verified on 21 September, reaches 10.1% on ARC-AGI-3 and 61.4% on ARC-AGI-2; Grok 5 has been announced for months with no verifiable date. <strong>DeepSeek V4.1 Flash</strong>, verified on 10 September, scores 3.3% and 72.9%, but it belongs to a family released as open weights, which makes it the best downloadable model in the table on ARC-AGI-2.</p>
<p><strong>Mistral Large 4</strong> is a different case: no score yet, but a promise. Announced on 6 October under the nickname “Le Chonk,” the model has 1 trillion parameters, was trained on around 4,000 Nvidia GPUs, “two to three times less than our Chinese competitors” according to Pierre Stock, Mistral’s VP of science, and its weights are due to be released within three weeks after safety testing (<a href="https://techcrunch.com/2026/10/06/mistrals-new-1t-model-aims-to-leapfrog-closed-and-open-rivals/" rel="noopener noreferrer nofollow">TechCrunch, 6 October</a>). Emmanuel Macron sees it as “a third way in AI.” Verification will come the day ARC Prize publishes its card.</p>
<h2 id="h-the-other-road-intelligence-without-a-data-center">The other road: intelligence without a data center&nbsp;</h2>
<p><strong>An object of a different nature, to be compared with care.</strong> Neuraxon is not a large language model. It is a network of simple units that coordinate and specialize, inspired by Marvin Minsky’s “society of mind,” which the Aigarth evolutionary framework improves from the networks generated by the Qubic network’s miners. The compute does not come from a data center but from <strong>useful proof of work</strong>, which redirects miners’ power toward training, a mechanism we detailed in <a href="https://www.cointribune.com/en/qubic-turning-every-cpu-cycle-into-real-world-value/" rel="noopener noreferrer nofollow">Qubic: Turning Every CPU Cycle Into Real-World Value</a>.</p>
<p>On 22 September, Qubic published a score of <strong>0.38 on ARC-AGI-3</strong>, up from 0.13 at the start, after 141 evolutionary improvement cycles and 450,000 attempts, on ordinary processors, with no GPU and no embedded language model (<a href="https://qubic.org/blog-detail/qubic-all-hands-recap-september-2026-neuraxon-arc-agi-3" rel="noopener noreferrer nofollow">Qubic recap</a>). In its 26 September analysis, the project’s science team is itself the first to set the limits: the evaluation runs on <em>“a few CPUs,”</em> and a distributed network remains “<em>much less capable at language tasks</em>” than a model trained on 100,000 GPUs (<a href="https://qubic.org/blog-detail/openai-agi-is-here-gpt-6-astra-neuraxon-intelligence-academy-vol-13" rel="noopener noreferrer nofollow">Qubic, Neuraxon Intelligence Academy vol. 13</a>).</p>
<p>Three reading precautions.</p>
<p>The 0.38 is not a score verified by ARC Prize and its protocol is not that of the semi-private set: it cannot be lined up against Astra’s 99.9%. There is no Neuraxon API comparable to ChatGPT or Claude. And Qubic’s thesis is not “<em>we are ahead,</em>” but “<em>AGI is an emergent, distributed and verifiable organization, not a model crossing a threshold.</em>” That is a position on governance as much as on technique, and it is on that criterion that the project stands apart from the table: open reference code, compute supplied by thousands of machines, results presented at conferences rather than in keynotes.</p>
<h2 id="h-what-the-ranking-does-not-say">What the ranking does not say</h2>
<p>A saturated test stops measuring. ARC-AGI-3 was designed to resist the models of 2025; two systems now clear it at over 97%, and ARC Prize warns that this does not prove AGI. Comparing different harnesses is like comparing runners with and without a tailwind: the same Astra scores 62.7% or 99.9% depending on whether its reasoning state is kept.</p>
<p>Cost almost never appears in press releases. Evaluating Astra cost up to $50,000 in API calls, while the 500 humans in the reference group were paid about $12.78 per game attempted. And OpenAI itself acknowledges that Astra is harder to monitor in control-evasion tests, a step back the company calls serious. Amelia Glaese, OpenAI’s VP of research: “<em>When models can do more things autonomously, we have to be able to trust them more.</em>“</p>
<p>The next dates are set. Mistral is due to release the Large 4 weights by the end of October; Google is widening Gemini 4 Argon to developers; Anthropic has not announced an ARC-AGI-3 run for Fable 5.1; Qubic will present the state of its useful proof of work in a keynote at the TUM Blockchain Conference in Munich on 29 October. The table will be updated with each card verified by ARC Prize.</p>

AI Comparison 2026: Who Is Ahead?

Table of Contents

Cointribune EN

Other articles published on Oct 25, 2026

Join our free newsletter for daily crypto updates!