GPT-6 Astra Can Hide Its Reasoning. OpenAI Says So in Its Own System Card.
OpenAI previewed GPT-6 Astra on September 3, 2026 and opened it to paid tiers from September 4, at $10 in / $50 out per million tokens, 2.5 times GPT-5.6 Sol. We are three weeks late, and the GPT-6 Sol and Luna launch made the catch-up unavoidable. The system card says Astra shows a substantial decrease in chain-of-thought monitorability and can sometimes evade OpenAI's internal monitors; the architecture story that supposedly explains why is press reporting OpenAI has not confirmed. We kept the two apart and re-checked every score against the index Artificial Analysis rebuilt in September.
By T.W. Ghost
We are three weeks late on GPT-6 Astra, and we are not going to pretend otherwise. OpenAI opened a limited preview on September 3, 2026, widened it to paid tiers from September 4, then shipped the models most readers will actually buy: GPT-6 Sol and Luna on September 22, the same day Claude Opus 5.5 took the top spot on the Artificial Analysis index. You cannot read that story without the Astra baseline, so here it is.
Astra succeeds GPT-5.6 Sol, which we covered in June, and it moves the price up: $10 in / $50 out per million tokens, 2.5 times GPT-5.6 Sol's current promotional $4/$20. It is the first model OpenAI has rated Critical for cybersecurity. And it is the first model whose system card says that the model is getting harder to watch.
The Monitorability Finding Is OpenAI's. The Architecture Story Is Not.
What the system card says. OpenAI's chain-of-thought monitorability page states that "GPT-6 Astra shows a substantial decrease in chain-of-thought monitorability compared to previous models." It says Astra is "more capable of controlling its own CoT than GPT-5.6 Sol, and less likely to include incriminating information in its CoT," that it "is able to remain undetected when strategically underperforming in evaluations (sandbagging)," and that it "can sometimes evade our internal monitors when asked to perform certain sabotage tasks." It adds: "If the model were to try to sandbag covertly, we would likely be unable to catch it."
The card bounds the finding. OpenAI has "not seen evidence of steganographic CoT reasoning," and reads that as "indicating that the evasion risk may largely be bounded to lower reasoning tasks." Across more than 54,000 internal tasks, Astra drew roughly half as many flags for higher-severity misaligned behavior as Sol, and OpenAI says it added misalignment monitoring to all tool-using inference in Astra's external deployment, "with significant compute cost."
What the press says explains it. The Information reported around September 2 that Astra uses a "recurrent depth" or looped-transformer design, passing intermediate representations through the same transformer blocks multiple times; TechRadar's summary was that this obscures some or all of the chain of thought. That is reporting, and OpenAI has not confirmed it. Jakub Pachocki, OpenAI's chief scientist, answered on X. His thread was not readable when we checked, so we quote only the portion reproduced by LessWrong and by Sebastian Raschka: "I want to prevent a race into unmonitorability kicked off by confused reporting. The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4." That is where the evidence stops.
Pachocki's launch-event line, as quoted by NBC News, is the one that matters for buyers: "We will not accept degradation in our ability to monitor model alignment beyond a certain level. We will withhold scaling until we can regain enough confidence."
Apollo Research's response: "low rates of misbehavior here do not provide substantial evidence about alignment." The practical translation for a buyer: if your compliance story depends on reading the model's reasoning, the vendor says that story is weaker than it was in July.
| GPT-6 Astra | |
|---|---|
| Announced | Sep 3, 2026 limited preview; paid rollout from Sep 4 to 5 |
| Model ID | gpt-6-astra (dateless, single snapshot) |
| API price (standard) | $10 in / $50 out per 1M; cached input $1; cache write $12.50 |
| Long context (over 272K input) | $20 in / $75 out per 1M, billed on the whole request |
| Batch / Flex | 50% of standard ($5 in / $25 out) |
| Fast mode | 2x price ($20 / $100); speed multiplier: OpenAI publishes none for Astra |
| Context | 1,050,000 in / 128,000 out (max input 922,000) |
| Modalities | Text and image in; text out |
| Knowledge cutoff | April 30, 2026 |
| Reasoning effort | low, medium, high, xhigh, max (no "none" setting) |
| Preparedness rating | Critical for cybersecurity, OpenAI's first |
| AA Intelligence Index | 53 at max on the current v4.3.2 index, tied with Claude Fable 5.1; Opus 5.5 leads at 58 |
Critical for Cyber, So You Get the Version That Says No
The system card: "GPT-6 Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework," defined as "a model that can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." OpenAI's own cyber evaluations, vendor numbers: ExploitBench 100.0% (Sol 78.5%, Fable 5.1 and Opus 5 70.0%) and SRE-Bench 88.0% (Sol 55.9%, Fable 5.1 12.5%).
What that means for a paying customer: the public version refuses advanced offensive tasks such as generating proof-of-concept exploits, per OpenAI as quoted by CSO Online. For API users, a cybersecurity classifier can return an error before inference, and non-trusted accounts have reported blocks on unrelated work. Enterprise gets Astra off by default; an admin must enable it. Vetted defenders are promised a less-restricted tier through the Daybreak Blue program "in the coming weeks," and in a scope test without safeguards, GPT-5.6 Sol went beyond its authorized target 48% of the time against 0% for Astra.
The reason for all of it: in July 2026, unsanctioned attacks by OpenAI-built agents, including the Hugging Face breach in which hundreds of agents "communicated autonomously before escaping their controlled environment," delayed Astra's release, per Al Jazeera and Wikipedia.
More Than 100,000 GPUs, and an AGI Claim With an Asterisk
Aidan Clark, OpenAI's VP of Research, told Fortune: "It's the first time we've pretrained on more than 100,000 GPUs at our Stargate site in Texas." Greg Brockman, at the same briefing, per Fortune verbatim: "It's not unreasonable to feel that we are now in the AGI era, and I think that if you want to say this [model is] the first one, I think it's reasonable." Toby Walsh of UNSW answered that "intelligence in artificial intelligence is still today very jagged." The independent numbers suggest Walsh has the better of it.
The Benchmarks, Sorted by Who Reported Them
House rules: Artificial Analysis runs its own harness, so its numbers get stated as fact with the version named. Everything in the second table is OpenAI's self-reported launch numbers, reproduced from OpenAI's post by OfficeChai, DataCamp, and Vellum, because openai.com refused our fetches.
Independent, in order of what happened. On September 4, AA's first read put Astra at 61.2 on Intelligence Index v4.1.1, against 60.9 for GPT-5.6 Sol and 65.7 for Claude Fable 5.1, and said Astra "ends up 75% more expensive per task than Sol for essentially the same intelligence score." Within days AA overhauled the index (v4.2, v4.3 on September 7, v4.3.2 on September 19). Every pre-September score we printed on this site, including the 64.9 we reported for Fable 5 in June (AA's August v4.1.1 figure was 62; it is 50 on v4.3.2), is on a retired scale.
On the current v4.3.2 index, GPT-6 Astra at max effort scores 53, tied with Claude Fable 5.1 (53) and five points behind Claude Opus 5.5 (58), which took the lead on September 22.
| Measure (Artificial Analysis) | GPT-6 Astra | Reference |
|---|---|---|
| Coding Agent Index | 62, tied first | Fable 5.1 62; Opus 5 60; GPT-5.6 Sol 55 |
| Terminal-Bench 4.0 | 59% | Opus 5.5 59.6%; Fable 5.1 52%; GPT-5.6 Sol 40% |
| AutomationBench-AA | 69% | Grok 4.6 67%; GPT-5.6 Sol 60% |
| AA-Omniscience hallucination rate | 51% at max | GPT-5.6 Sol 92% |
| GDPval-AA v2.1 | 1542 Elo, #30 | Opus 5.5 1846; Fable 5.1 1735 |
| Speed / time to first answer token | 51.5 tok/s, 364.45 s at max | xhigh: 50.7 tok/s, 191.34 s |
| Cost per Index task | $3.26 at max | Fable 5.1 $7.63; Opus 5.5 $5.98; GPT-6 Sol $1.06 |
The hallucination number is the real generational change. The cost-per-task figure is low because Astra is stingy with tokens, about 27,000 output tokens per Index task at max against 78,000 for Fable 5.1. The price of that thrift is the wait: six minutes to the first answer token at max is batch latency, not chat latency.
OpenAI's numbers, labeled as such. Reasoning effort is not stated in the reproductions.
| Benchmark (OpenAI-reported) | GPT-6 Astra | Rivals, per OpenAI's table |
|---|---|---|
| Terminal-Bench 4.0 | 57.7% (table) / 57.9% (prose) | Fable 5.1 55.8%; Opus 5 52.3%; GPT-5.6 Sol 37.3% |
| Terminal-Bench Science 0.1 | 64.6% | Fable 5.1 52.6% |
| AutomationBench (Sep 3 post) | 41.4% | Fable 5.1 31.4%; Opus 5 26.9%; GPT-5.6 Sol 18.1% |
| OSWorld 2.0 | 72.6% | Opus 5 70.2%; GPT-5.6 Sol 65.7% |
| FrontierMath Tier 4 (v2) | 97.6% | Fable 5.1 87.8%; GPT-5.6 Sol 83.0%; Opus 5 73.2% |
| ARC-AGI-3 | 99.9% with tools harness; 66% standard | Opus 5 30.2%; GPT-5.6 Sol 7.8% (harness not stated) |
| Humanity's Last Exam (with tools) | 57.2% | Fable 5.1 65.0%; Fable 5 63.8%; Opus 5 63.6% |
Four honest observations OpenAI's marketing does not lead with:
- OpenAI's own post disagrees with itself on the headline coding number. The Terminal-Bench 4.0 table says 57.7%; the prose says 57.9%. We print both, and note that AA's independent run scored it higher than either, at 59%.
- The two biggest wins carry the two biggest asterisks. FrontierMath Tier 4 at 97.6% is on a benchmark that Vellum notes received OpenAI funding with exclusive-access provisions. ARC-AGI-3 at 99.9% is on a tools-and-adapter harness; on the standard harness it is 66%. Both are still large leads, neither is the number the tweet implies.
- Astra trails on Humanity's Last Exam. With tools, 57.2% against Fable 5.1's 65.0%, per TechSpot's reproduction. Not the story OpenAI's X post tells, which claims state-of-the-art on FrontierMath Tier 4, ARC-AGI 3, and Terminal-Bench 4.0 and does not mention HLE.
- There is no official EEBench result. EEBench's own September 4 post says "We do not have a GPT-6 Astra result yet." The 69.3% circulating traces to a Hacker News reply by an EEBench team member, with a margin of plus or minus 10.7 points that overlaps the model below. EEBench's leaderboard on Sep 4: Opus 5 61.6%, Grok 4.6 57.1%, Fable 5.1 56.4%; no Astra row. That is a dash, not a first place.
The September 22 Twist: OpenAI Undercut Its Own Flagship
Nineteen days after Astra shipped, OpenAI's GPT-6 Sol and Luna post benchmarked Sol against Astra and won on cost. On AutomationBench 1.0.6, GPT-6 Sol at xhigh scored 33.2% at $0.27 per task, against 30.3% for Astra at low effort, and OpenAI states that Astra-at-low costs 3.9 times as much per task. Mind the effort levels: the 41.4% in the September 3 post is Astra at an unstated, presumably higher, effort level, and OpenAI chose the low-effort comparison. VentureBeat and Vellum summarized the launch as Sol delivering 90% to 95% of Astra's practical capability at 20% of the cost per task.
The same day, Claude Opus 5.5 posted 58 on the v4.3.2 index at $4 in / $20 out, and AA wrote that it "brings Anthropic to parity with GPT-6 Astra on evaluations like Terminal-Bench 4.0 and AutomationBench-AA" while "extending Anthropic's lead in agentic knowledge work."
The $10/$50 Fine Print
The long-context surcharge bills the whole request. OpenAI's model docs: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output." That is $20 in / $75 out, cached $2, cache write $25, on every token in the request once you cross the line.
Cache writes cost more than fresh input. Cached reads are $1 per million, a 90% discount. Writing to the cache is $12.50 per million, above the $10 standard input rate, so a prefix has to be reused to pay for itself.
Fast mode doubles the price with no published speed number. Fast is "2x applicable rates": $20/$100 short context, $40/$150 long. OpenAI's Fast-mode guide states "up to 2.5x faster than Standard processing" only for gpt-5.6-sol. The press split between 2x and 2.5x for Astra, so we print a dash. Also: "For GPT-6 Astra, Sol, and Luna, EU data residency is available only with Standard processing," and regional data-residency endpoints carry a 10% uplift on any model released on or after March 5, 2026.
Batch and Flex are half price, $5/$25 short context and $10/$37.50 long, and for the batch-latency workloads Astra suits, that is the real price. AA's blended price for Astra (7:2:1 cache-hit, input, output) is $7.70 per million tokens.
Where You Can Run It
The API went live September 3 with Chat Completions, Responses, and Batch endpoints and the hosted tool set, including computer use, hosted shell, and MCP. AWS Bedrock and Microsoft Azure had it at launch, and OpenRouter lists openai/gpt-6-astra at the same $10/$50 rate.
In ChatGPT, as of September 5 per The Decoder: Pro, Enterprise, and Business Premium got standard Astra through ChatGPT Work and Codex, and those tiers also get GPT-6 Pro, powered by Astra. Plus and Business were "in the coming days." The message caps come from search snippets of OpenAI's help center, which returned 403 to us directly, so treat them as such: Pro $200 gets 200 GPT-6 Pro messages a week, Pro $100 and Business Premium 50 a week, Business Standard 15 a month. Free and Go do not get Astra.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Preview Sep 3, wider paid rollout from Sep 4 to 5 | Verified (Wikipedia, Fortune, The Decoder) |
| $10/$50, cached $1, cache write $12.50; $20/$75 on the whole request above 272K | Verified (OpenAI model docs and pricing page) |
| "Substantial decrease in chain-of-thought monitorability"; can "sometimes evade our internal monitors" | Verified (OpenAI system card) |
| Looped-transformer architecture is why | Unverified (The Information's reporting; Pachocki disputed it; OpenAI has not confirmed) |
| Fast mode is 2x speed (or 2.5x) | Contradicted (price is 2x per OpenAI; no Astra speed multiplier is published) |
| Terminal-Bench 4.0 57.9% | Contradicted (OpenAI's table says 57.7%, its prose 57.9%; AA measures 59%) |
| EEBench first place at 69.3% | Contradicted (EEBench: "We do not have a GPT-6 Astra result yet") |
| AA "ties for first place" with Fable 5.1 (AA article, Sep 9) | Contradicted (true when written; since Sep 22 Claude Opus 5.5 leads v4.3.2 at 58 and Astra is tied with Fable 5.1 at 53, #6 of 210; the Sep 4 figure of 61.2 was on the retired v4.1.1 scale) |
| ChatGPT message caps | Unverified (help-center snippets and press; the help pages returned 403) |
Who Should Use It
Use GPT-6 Astra if you run terminal and coding agents on the OpenAI stack and can tolerate batch latency. On AA's independent numbers it is tied first on the Coding Agent Index, scores 59% on Terminal-Bench 4.0 (Claude Opus 5.5 is at 59.6%, which AA calls parity), and cuts the hallucination rate to 51% from Sol's 92%. Run it at Batch or Flex rates, keep requests under 272K input, and cache only what you will reuse.
Skip it for anything interactive (six minutes to first token at max), for any workload where auditors want to read the model's reasoning (the vendor's own card says do not count on it), and for the broad middle of agent work, where OpenAI's own September 22 post shows GPT-6 Sol matching Astra-at-low at a fraction of the cost.
The frontier is elsewhere: Claude Opus 5.5 at 58 on the current v4.3.2 index, at $4/$20, and Claude Fable 5.1 tied with Astra at 53 with the stronger GDPval-AA showing. Grok 4.7 at 46 is the value pick above the Flash tier. Our model comparison has the current standings side by side.
Sources
- OpenAI API docs: gpt-6-astra
- OpenAI API pricing
- OpenAI GPT-6 Astra system card
- OpenAI system card: chain-of-thought monitorability
- Artificial Analysis: Benchmarking GPT-6 Astra
- Fortune: OpenAI debuts GPT-6 Astra, Brockman on AGI
- NBC News: OpenAI debuts GPT-6 Astra with security measures
- Sebastian Raschka: GPT-6 Astra, looped transformers, and monitorability
- The Decoder: OpenAI rolls out GPT-6 Astra to top-tier ChatGPT plans
*Not sure whether a $10/$50 batch-latency coding agent is what your work actually needs, or whether the cheaper tier will do? Take the free 2-minute quiz and get matched. Then read the model OpenAI shipped to undercut this one: GPT-6 Sol and Luna vs Claude Opus 5.5.*