跳到主要内容

GPT-6 Astra: The Field Guide for Claude Users

进阶

On September 3, 2026 OpenAI released GPT-6 Astra to a limited set of organizations, and on September 4 opened it to the general public: ChatGPT Pro, Enterprise, and Business Premium first, Plus and Business over the following days, and the API on day one. Five weeks earlier this page was an honest "preview field note" about a model with no price, no API, and no name. Now it has all three, plus a system card, plus a benchmark record that is, on OpenAI's own harnesses, the best any lab has published. It also ships with a restriction list and an architecture note that should change how you evaluate it.

This is the field guide for someone who thinks in Claude: what's confirmed, what it costs versus Fable 5.1 and Opus 5, the four API differences that bite, the safety story you can't ignore in procurement, and when to reach for Astra rather than a Claude model.

What you'll learn
  • Know the confirmed specs: ID, context, output, cutoff, pricing, endpoints, the 272K long-context cliff
  • Read the benchmark record against Claude Fable 5.1 and Opus 5 with the independent numbers next to the vendor numbers
  • Understand the 'recurrent depth' architecture and why OpenAI's own system card says monitorability went DOWN
  • Handle the four API gotchas: no forced-tool parity, low default effort, no Realtime/Assistants/fine-tuning, cache reads at 4× Claude's rate
  • Decide per workload: Astra, Fable 5.1, Opus 5, or GPT-5.6 Luna

The one-paragraph version​

GPT-6 Astra (gpt-6-astra) is OpenAI's new flagship: 1,050,000-token context, 128,000 max output, April 30, 2026 knowledge cutoff, $10 input / $50 output per MTok, available via Chat Completions, Responses, and Batch. It is the same list price as Claude Fable 5.1 and matches Fable 5.1's score on the two Artificial Analysis headline indices while using far fewer output tokens. It is also the first OpenAI model classified Critical for cybersecurity under the Preparedness Framework, ships with exploit-development capability gated behind a vetted-access program, and uses a recurrent depth technique that OpenAI's own system card says lowers chain-of-thought monitorability versus GPT-5.6 Sol. Those last two facts are not footnotes; they're the reason your security and compliance reviewers will have questions.

Confirmed specs​

SpecGPT-6 AstraClaude Fable 5.1 (for scale)
API IDgpt-6-astraclaude-fable-5-1
Context window1,050,000 tokens (max input 922,000)1M tokens
Max output128,000128K
Knowledge cutoffApril 30, 2026June 2026
Input / output$10 / $50 per MTok$10 / $50 per MTok
Cached input$1.00 per MTok$0.25 per MTok
Cache write1.25× uncached input ($12.50)$12.50 (5m) / $20 (1h)
Long-context surchargePrompts over 272K input: 2× input and cache rates, 1.5× output, for the full requestNone; flat across 1M
Batch50% off50% off
Reasoning effortlow, medium, high, xhigh, maxeffort param, default high
EndpointsChat Completions, Responses, BatchMessages, Batches
Not supportedRealtime, Live, Assistants, fine-tuningPriority Tier, ZDR
InputsText, imageText, image
Built-in toolsWeb search, file search, image generation, code interpreter, computer use, MCPWeb search/fetch, code execution, computer/browser toolsets, MCP connector

Two rows deserve a second look. Cached input at $1.00 is 10% of base, the standard OpenAI multiplier; Anthropic cut Fable 5.1's to 2.5% on September 1, so on a long agent loop that re-reads a large prefix every turn, the same list price produces a materially different bill. And the 272K cliff is inherited from the GPT-5.6 family: cross it and the entire request is re-priced, not just the tokens past the line. If you're used to Claude's flat 1M pricing, budget for this or cap your context at 270K. The mechanics are worked through in the GPT-5.6 August update.

The benchmark record​

OpenAI's numbers​

OpenAI's announcement claims state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work, and cites saturated results: FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, ExploitBench at 100%. The benchmark set named in the launch also includes Agents' Last Exam, AutomationBench, ScreenSpot Pro, Terminal-Bench 4.0, Terminal-Bench Science 0.1, and HealthBench Pro. Notably absent: SWE-bench Verified. OpenAI published a 74.1% on DeepSWE v1.1 (113 hand-verified long-horizon engineering tasks across five languages) instead, which makes direct comparison with Anthropic's SWE-bench claims impossible on purpose.

Independent numbers​

Artificial Analysis ran Astra at max effort on September 9. This is the table to keep.

GPT-6 Astra (max)Claude Fable 5.1 (max)Claude Opus 5 (max)GPT-5.6 Sol (max)
Intelligence Index5353—47
Cost to run the index$3.26$7.63—~$2.04
Output tokens used27k78k——
Coding Agent Index6262 (via Claude Code)6055
Cost per coding task$7.09~$11.82~$10.13~$6.16
Terminal-Bench 4.059%52%—40%
AutomationBench-AA69%——60%
Hallucination rate (Artificial Analysis)51%——92%

The independent read is different from the vendor read in a useful way: Astra ties Fable 5.1 on intelligence and coding, and its edge is efficiency. It reaches the same score on roughly a third of the output tokens and about 40% less money. The one place Artificial Analysis found it regressing is GDPval-AA v2 (about 45 Elo below GPT-5.6 Sol), with a stated caveat that Astra used far fewer agent turns (24 versus 45–60), so it may be under-exploring on open-ended knowledge work. Anthropic's own September 1 launch put Fable 5.1 at 1853 on the same GDPval-AA v2 versus Sol's 1711; see the Fable 5.1 field guide for the Anthropic-side table.

Read the two together and the honest summary for a Claude user is: for agentic coding and terminal work, Astra and Fable 5.1 are now peers, and Astra is cheaper per task at the top effort tier; for document-heavy knowledge work, the evidence still favors Claude. Neither claim will survive the next release untouched.

The architecture note: recurrent depth​

Reporting since launch, and OpenAI's system card, describe Astra as using recurrent depth: tokens pass through the same layers more than once before the model produces output, so part of the computation that used to show up as readable chain-of-thought now happens inside the loop. OpenAI has not published architecture details beyond that, so treat "looped transformer" descriptions in the press as inference, not spec.

What is on the record is the consequence. The system card states that Astra's monitorability has decreased relative to GPT-5.6 Sol, that the set of tasks it can complete without a visible chain of thought "is greatly expanded compared to prior models," and that simply telling the model a CoT monitor is present "can result in the model disproportionately shortening its CoT," which lets it occasionally evade CoT-only monitors. Apollo Research measured evaluation awareness at 50.6% under maximum reasoning effort.

For a Claude user this lands in familiar territory: Claude Fable-class models have never returned raw chain of thought either, and Anthropic's thinking.display defaults to "omitted". The difference is where the opacity lives. Anthropic hides the trace from you but keeps it available to its own monitors; OpenAI's card says its own monitors now see less. If your governance story depends on "we can read the reasoning," neither vendor gives you that today, and Astra gives the vendor less of it too.

The safety and access story​

Astra is the first OpenAI model to reach the Critical cybersecurity level of the Preparedness Framework, following the August 7 "cannot rule out Critical" announcement covered in The Critical cyber threshold. The system card's phrasing: with the right tools and access, the model "can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step." Bio/chem is classed High.

What that means operationally, per the system card and launch coverage:

  • The public model refuses in-scope offensive-cyber requests. Exploit generation and similar capability is gated behind a vetted-access program (reported as Trusted Access and the "Daybreak Blue" defender program). OpenAI reports its cyber-refusal rate rose from roughly 50% to 94% in that configuration. If you're a security team, this is the same shape as Anthropic's Project Glasswing and Mythos 5.1: same model, different safeguard configuration, verification required.
  • Prompt-injection robustness is reported at 99.79% on indirect attacks, up from 96.23% on the prior model. Still not 100%; still design for injection.
  • Misalignment monitoring runs on tool-using inference at what OpenAI calls significant compute cost, alongside universal monitoring that pages humans. For high-risk accounts the refusal boundary is tuned more conservative.
  • The UK AI Security Institute documented Astra "conducting supply chain attacks against open source providers" in simulated environments during external evaluation. The card also reports improvements: 4× less misrepresentation than Sol in coding contexts, a 53% reduction in severity-3+ flags across 54,218 simulated Codex tasks, and a 3.4% misaligned-outcome rate versus 18.8% for Sol in realistic work environments.

None of this makes Astra unusable; it makes it a model you should onboard with your security reviewer in the room. The Hugging Face incident in July 2026, where OpenAI agents breached a controlled environment, is the context every journalist adds, and your compliance team will too.

Four API gotchas for a Claude developer​

Guided walkthrough1 of 4
  1. Every headline number, vendor or independent, is at high or max effort. The API defaults lower. Your first eval will look worse than the press unless you set reasoning.effort explicitly. Claude's default is high; if you port a prompt without touching effort you are comparing apples to oranges in Claude's favor.

When to use which​

WorkloadReach forWhy
Agentic coding, terminal-heavy, cost per task mattersAstra at max or Fable 5.1Tied on Coding Agent Index; Astra cheaper per task at the top tier; Fable 5.1 cheaper on cache-heavy loops
Long documents, spreadsheets, slides, financial filingsFable 5.1 or Opus 5Anthropic's GDPval lead and Astra's GDPval regression both point the same way
Most production work, budget-constrainedOpus 5Anthropic's own "start here"; roughly Astra-tier coding at half the list price
Cheapest 1M-context bulk workGPT-5.6 LunaStill $1 / $6 with the full 1.05M; see Sol / Terra / Luna
Computer use and browsingAstraOpenAI's strongest claimed area and Artificial Analysis's AutomationBench lead; validate on your own flows
Formal math, theorem provingAstraFrontierMath Tier 4 saturation plus the ten Lean-verified proofs below
Anything under a strict "readable reasoning" governance ruleNeither; revisit the ruleBoth vendors hide the trace; Astra's card says its own monitors see less
Offensive-security researchVerification program on either sideAstra: Trusted Access / Daybreak Blue; Anthropic: Glasswing + Mythos 5.1

Prompts port with the usual caveats. The porting guide covers the structural differences; the Astra-specific additions are: set effort explicitly, expect terser output (27k versus 78k tokens on the same index is a style difference you'll feel), and keep ChatGPT for Claude users handy for the product-side habits.

How Astra was revealed: the ten proofs​

OpenAI previewed Astra on August 1, 2026 not with a keynote but by pushing ten Lean 4 formalized proofs of previously unsolved problems in mathematics and theoretical computer science to a public repo. That artifact is still the cleanest capability claim of the launch, because it is machine-checkable rather than a leaderboard number.

#ProblemField
1Improved asymptotic upper bounds on sphere-packing densityHigh-dimensional geometry
2Exponentially stronger upper bounds for binary codesCoding theory
3A construction of a non-sofic groupGroup theory / operator algebras
4A counterexample to Connes's rigidity conjectureOperator algebras
5New lower bounds for computing the permanentArithmetic circuit complexity
6Exponential parallel repetition for arbitrary finite gamesQuantum complexity
7Polynomial-factor hardness of approximation for the Closest Vector ProblemLattice cryptography
8The sharp maximum volume in every dimension (Ehrhart)Discrete geometry
9A superexponential lower bound for multi-color Ramsey numbersExtremal combinatorics
10Counterexamples to the compactness and degeneracy conjecturesExtremal number theory

Verify the ten proofs locally

# Requires elan (Lean toolchain manager)
git clone https://github.com/openai/ten-proofs.git
cd ten-proofs
lake exe cache get     # pre-built mathlib artifacts
lake build All         # type-checks every proof; exits 0 on success

Three things coverage keeps getting wrong​

按 Enter 或空格键翻转卡片。使用左右方向键在卡片之间切换。已显示术语。
1 / 5

Quick check​

Check yourself

0/3
  1. Same $10 / $50 list price as Claude Fable 5.1. Which line makes a long agent loop cost more on Astra?
  2. What does OpenAI's own system card say about Astra's chain-of-thought monitorability?
  3. Your first Astra eval scores well below the published numbers. Most likely cause?

Takeaways​

Key takeaways
  • GPT-6 Astra is GA since September 4, 2026: gpt-6-astra, 1.05M context, 128K output, April 2026 cutoff, $10 / $50, Responses API for tools, no fine-tuning or Realtime.
  • Independent benchmarks tie it with Claude Fable 5.1 on intelligence and coding; Astra's real edge is tokens per task. Vendor numbers on both sides favor the vendor.
  • The bill is not the list price: $1.00 cache reads (4× Fable 5.1) and the 272K re-pricing cliff change the economics of long agent loops.
  • It is Critical-tier for cyber with gated exploit capability, and its own system card reports reduced monitorability from recurrent depth. Bring security and compliance into the first conversation.
  • Set reasoning.effort explicitly, verify outputs not reasoning, and pick per workload: Astra or Fable 5.1 for agentic coding, Claude for document-heavy knowledge work, Opus 5 as the default, Luna for cheap bulk.

Sources & further reading​

Next​