AI Brief, 4 September 2026: one model, one benchmark, 37 points apart

OpenAI released GPT-6 Astra on 3 September, and the most useful thing published about it did not come from OpenAI. ARC Prize evaluated the model the day before launch and posted results the same day, and those results contain a number that should change how the next twelve months of benchmark reporting are read. On ARC-AGI-3 Semi-Private, Astra scores 62.71% under ARC's neutral harness and 99.95% under a harness that uses OpenAI's own context-management features. Same model, same tasks, same evaluator, on the same day. The gap between the two conditions is larger than the gap between Astra and every model that came before it.

The model itself is real but narrowly available. It is not generally available: the launch post says it is rolling out to a limited set of organisations, and OpenAI's API documentation is narrower still, describing a rollout to enterprises in a Trusted Access Program with broader ChatGPT and API access "in the coming days". It is priced at $10 per million input tokens and $50 per million output, two and a half times GPT-5.6 Sol. It carries a 1,050,000-token context window. And it ships still designated Critical for cybersecurity under OpenAI's Preparedness Framework, the first model OpenAI has ever placed at that level, which this brief covered when the designation was announced on 1 September. OpenAI publishes no architecture description whatsoever: no parameter count, no training compute, no structural detail. Reports circulating that Astra uses a recurrent architecture are not supported by anything OpenAI has written.

The day's other confirmed event closes a story this brief left open on 27 August. Nvidia said on 3 September that it has agreed to acquire Hugging Face, and Jensen Huang's post names a figure to the dollar: $12,930,300,000. That week, Business Insider reported talks above $13 billion with no signed agreement while The Information reported an agreed deal at $12.9 billion, and this brief said the thing to watch was whether either company would say anything on the record. Both outlets turn out to have been right about different things. Nvidia's Form 8-K, filed 3 September, gives the agreement date as 2 September, six days after those reports.

• GPT-6 Astra shipped 3 September to a gated enterprise cohort, at $10/$50 per million tokens, 1.05M context, still designated Critical for cyber.
• ARC Prize measured 62.71% on ARC-AGI-3 with its own harness and 99.95% with a provider-supplied one, and says it will now report both, adding that it is "not claiming that it is AGI".
• Artificial Analysis independently scores Astra at 61.2 on its Intelligence Index, five points behind Claude Fable 5.1 at 65.7 — a ranking OpenAI reproduces in its own launch table.
• Nvidia's 8-K splits the $12.93 billion into roughly $11.9 billion to shareholders plus up to $1.0 billion in equity retention, with closing expected in the first half of 2027.
• The commitment to keep the Hub open to rival silicon is in the 8-K, phrased as a corporate commitment; the merger agreement itself was not filed, so nothing about it is publicly enforceable.
• Hugging Face has published no statement of its own. The only remarks from its side are Clément Delangue's live on CNBC.

Astra ships gated, and stays Critical #

What OpenAI published is a large benchmark table and a 210 KB system card, all of it the lab's own measurements run in its research environment rather than production ChatGPT. Terminal-Bench 4.0 comes in at 57.9% against Claude Fable 5.1's 55.8% and Sol's 37.3%; FrontierMath Tier 4 at 97.6%; MRCR 8-needle retrieval at 96.3% in the 512K–1M band against Sol's 73.8%. On the cyber evaluations the margins are wider: ExploitBench 100.0% against Sol's 78.5%, SRE-Bench 88.0% against 55.9%.

Two rows in that table are worth more than the wins. OpenAI reproduces Artificial Analysis's Intelligence Index and its own Coding Agent Index inside the launch post, and Astra places third on both; on Humanity's Last Exam with tools it scores 57.2% against 65.0% for Fable 5.1. A lab publishing third-party numbers that rank it behind a competitor, in its own launch material, is not the norm.

The safety position is more interesting than the capability one. The shipped model refuses to build proof-of-concept exploits; OpenAI reports a cyber jailbreak refusal rate of 91.5% against Sol's 59%, and says it runs misalignment monitoring on all tool-using inference in external deployment, at what it calls significant compute cost. On a scope-overrun evaluation built after the July Hugging Face intrusion, Astra scores 0% against Sol's 48%.

Against that, one result runs the other way and OpenAI publishes it. Astra is substantially better at controlling its own chain of thought: asked to keep reasoning within a 750–1,250 token band, it complies 60.9% of the time against Sol's 16.1% and GPT-5.5 Thinking's 1.7%. The system card reports that the model can sandbag without detection, sometimes evades internal monitors on sabotage tasks, and that some successful honeypot attacks contained zero chain-of-thought tokens at all — caught only because full-trajectory monitoring does not depend on the reasoning trace. Monitorability got worse as capability went up, and OpenAI attributes that to training rather than "any architectural changes".

One correction to the coverage: OpenAI never claims AGI. Every occurrence of the string on the launch page is a benchmark name. Headlines framing this as a lab declaring the AGI era describe the reception, not the document.

The same model, the same benchmark, 37 points apart #

ARC Prize's write-up, by Greg Kamradt, tested Astra on 2 September across twelve harness configurations and reports every one. The two that matter are defined on the results pages in a sentence each.

The Standard harness "enables a model to carry forward notes it chooses to keep with it throughout the environment". It is provider-neutral and minimal: between turns, the only thing that survives is text the model wrote down on purpose, which ARC can read. The Provider Adapter harness "preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work". That is OpenAI's own context management, and ARC does not see what is retained.

# Standard: the only thing that survives a turn is what the model chose to write.
notes = ""
for frame in game:
    reply = model(prompt=[SYSTEM, notes, render(frame)])   # fresh context every turn
    notes = reply.notes                                     # visible to the evaluator, lossy
    act(reply.action)

# Provider adapter: the provider carries the model's internal state forward.
state = None
for frame in game:
    reply = model(prompt=render(frame), carry=state)        # opaque continuation
    state = reply.state                                     # never inspected by the evaluator
    act(reply.action)

The effect is not a few points. It is the whole benchmark.

Under the Standard harness the effort dial does most of the work, taking the score from 17.45% at low effort to 62.71% at max. Under the Provider Adapter it does almost none: the score sits between 98.03% and 99.95% at every setting, and the highest figure comes from high effort, not max. Whatever the adapter preserves substitutes for the reasoning the effort dial buys.

Cost inverts the usual reading. The Standard max run cost $26,097.50 in total and scored 62.71%. The Provider Adapter high run cost $18,816.63 and scored 99.95%. That is $416 per point of ARC-AGI-3 under the neutral harness against $188 per point under the adapter — the better score was 28% cheaper. Across the 167 game-reasoning pairs both harnesses solved, the adapter used 49% fewer tokens and ran roughly 3.66 times faster.

ARC's other headline is about efficiency, and its denominator matters. Under the Provider Adapter, Astra at max effort "used fewer actions than the human baseline on 96.0% of levels" and 51.7% fewer actions per level on average. That baseline is the median action count among roughly 500 members of the general public who completed the level, not selected for puzzle ability, and the 96.0% is over levels Astra itself completed. For scale on the jump: Claude Opus 5 at high effort scored 30.16% on ARC-AGI-3, GPT-5.6 Sol at max 7.78%. ARC's post carries no disclosure of OpenAI funding or pre-release support, which is not the same as a statement that there was none.

Artificial Analysis published independently on 3 September as well, and its picture is more ordinary. Astra at max effort scores 61.2 on the Intelligence Index at $0.899 per task, against Claude Fable 5.1 at 65.7 and $1.955, Sol at 60.9 and $0.562, and Gemini 3.8 Flash at 58.7 and $0.339. Astra ties Sol while costing 75% more per task. Its measured hallucination rate is 51.34%, against 92% for Sol. No pre-release-evaluation disclosure appears in the article.

Nvidia's openness commitment lives in an 8-K #

Nvidia's blog post carries the price and the reassurance; the Form 8-K, signed by CFO Colette Kress, carries the terms. The $12.93 billion splits into approximately $11.9 billion payable to Hugging Face stockholders, subject to adjustment, plus an equity-based retention programme of up to approximately $1.0 billion for employees joining Nvidia. Coverage describing $12.9 billion going to shareholders is overstating it by a billion. The form of consideration — cash, stock, or a mix — is not disclosed anywhere. Closing is expected in the first half of 2027, subject to regulatory approvals that the filing does not enumerate.

The neutrality question has an answer, and it is worth reading exactly. The 8-K says Nvidia "has committed to" keeping the platform open "consistent with Hugging Face's existing practices", and that under that commitment Hugging Face "would continue to permit" uploads and downloads of models and datasets of users' choosing "and to support other silicon vendors". The blog adds that Nvidia compute "will not be required" to build or deploy through the Hub. AMD and Intel are never named. The filing was made under Item 8.01, Other Events, rather than Item 1.01, Material Definitive Agreement, and the merger agreement was not filed as an exhibit — so the covenants behind that commitment are not public, no duration is stated, and no third party is named as able to enforce it. It is a stated corporate commitment, which is a real thing, and not a demonstrated binding one.

Delangue told CNBC on 3 September that all three founders and the team are joining, with the goal of running an "independently neutral platform within the NVIDIA team". No governance structure, separate subsidiary or reporting line appears in any primary document. Huang said on the same programme that there were other bidders, unnamed, and $12.93 billion is what it took. Hugging Face's own blog and press channels carry nothing about the acquisition as this issue goes out, so its only statement on record is spoken.

One piece of history sharpens the price. TechCrunch, citing the Financial Times, reported on 27 August that Hugging Face turned down a $500 million Nvidia investment in late 2025 at a $7 billion valuation, because it did not want a dominant investor able to sway its decisions. That objection has now been resolved by acquisition.

Also notable #

• A two-person team quantised the recurrent half of a hybrid 27B model to 4 bits and lost half a point. arXiv 2609.04098, submitted 3 September by Sergii Kozyrev and Davyd Maiboroda of mnma.ai with no institutional affiliation, takes Qwen3.8-27B — 48 Gated DeltaNet layers to 16 full-attention layers — and pushes all 496 linear layers to NVFP4 with 4-bit weights and activations. Five-task average falls from 85.62 to 85.10; weights drop from 50.13 to 17.53 GiB. The argued mechanism is that the gates are parameterised in log space, so an 11.0% error in the GEMM feeding the forget gate collapses to 2.1% at the block output, while 0.1% noise injected directly into the gate produces 22% state error. Every number is the authors' own. The most practically useful part is a serving bug: a per-module calibration scale colliding with vLLM's fused Gated DeltaNet projections made a broken model look better on long-context perplexity, 6.86 against a true 10.84, while AIME accuracy quietly fell to 80.8.
• Google published WeatherNext 3 at 15:00 UTC on 3 September: hourly rather than six-hourly forecasts at 0.1° resolution, with a 43-page technical report. There are no open weights; forecast data goes out through BigQuery, Earth Engine and the Maps Platform Weather API. The skill claims are Google's own — up to 60% lower precipitation CRPS against IMERG at early lead times, roughly 10% average improvement on upper-level variables in week one against ECMWF's AIFS ENS v2. The "most accurate global weather model to date" line is attributed to independent live evaluations by Brightband, which the technical report does not cite.
• OpenAI committed $1 billion to Daybreak for frontline defenders on 3 September. It is a subsidy programme, not a model: discounted access to existing Daybreak cyber models plus training and support, targeted to be consumed within six months, US first, with an MS-ISAC pilot. OpenAI says roughly 2,000 organisations are already approved.
• GitHub is retiring four Copilot models on 2 October: Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7. Gemini 3.8 Flash went generally available in Copilot the same day, on introductory provider pricing through 31 December.
• The reported $35 billion Anthropic–Lambda compute deal still has no primary source. Neither company's newsroom carries it. It originated with The Information on 31 August and was matched by Bloomberg and the Wall Street Journal; the 3 September pickups add nothing. Lambda's own 27 August post about a $926 million term loan backing GPU deployment for "an investment-grade customer" does not name one.
• Thinking Machines is reported to be raising about $1 billion at a valuation of at least $40 billion, with Accel in talks to lead. The Information reported it and TechCrunch matched it with a separate source on 3 September. Nothing is closed and neither party commented.

What to watch #

• Which ARC-AGI-3 number becomes the quoted one. ARC will publish both harnesses; OpenAI's launch table already cites 99.9. Within a week the figure that appears in a slide with no harness named will tell you which convention won.
• Whether the two zero-days Astra found get disclosed. The 2 September issue flagged this as a coordinated-disclosure question with a public clock on it. OpenAI now says disclosure to maintainers is under way, with no date and no CVE identifiers, so the clock is still running.
• Whether Nvidia ever files the merger agreement. Filing under Item 8.01 with no exhibit keeps the openness covenants private. An amended filing, or a regulator's published undertaking, is what would turn a commitment into an obligation — and the first-half-2027 close leaves a long window for one.
• Whether Hugging Face says anything in its own words. A company whose entire value is neutrality has so far let its acquirer describe what neutrality will mean.
• Independent numbers for Hy4 preview, GLM-5.3 and DeepSeek-V4-Flash-Vision-Exp are missing for a fourth issue. Artificial Analysis added Astra within a day of launch and has still added none of these, so every figure for all three remains the vendor's own.