by OpenAI · GPT-6 family · best for hardest reasoning, math and long-running agent work
GPT-6 Astra is OpenAI's new top model, generally available in the API since 2026-09-03. OpenAI describes it as "our most capable model, built for the hardest end-to-end work" and places it above GPT-5.6 Sol as a new flagship tier, not as a named successor to any earlier model. It is a reasoning-only model (no none effort) with a 1.05M-token window, 128K output, and a reasoning ladder that runs from low to max. Early independent results put it at or near the top of the field: first on Epoch AI's Capabilities Index, the highest GPQA Diamond and FrontierMath scores in Epoch's runs, first on Arena's WebDev board, and state of the art on ARC-AGI-3. Human-preference chat rankings do not yet show a lead. The buyer's take: this is the model to test for the hardest reasoning, math and long-running agent work, at the same $10/$50 list price as Claude Fable 5.1 and 2.5x GPT-5.6 Sol.
configuration_update items that change reasoning effort mid-conversation without breaking the prompt cache. OpenAI's reasoning guide lists Astra as the only model that supports configuration updates.low, medium, high, xhigh and max. none is removed and returns HTTP 400. Pro mode is supported and bills at standard token rates.temperature, top_p and log probabilities are no longer accepted. Tool calling requires the Responses API.| Benchmark | Score | Source |
|---|---|---|
| LMArena Elo | 1480 | lmarena.ai 2026-09-13T00:00:00.000Z |
| GPQA Diamond | 95.8% | epoch.ai 2026-09-03T00:00:00.000Z |
| LMArena Coding Elo | 1543 | lmarena.ai 2026-09-13T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The strongest measured model OpenAI has shipped, now on Azure and Bedrock too, but it is a 2.5x-Sol escalation tier, not a default.”
Astra gives OpenAI a credible claim to the top of the frontier again, with independent evaluators (Epoch, ARC Prize, Arena WebDev) rather than only OpenAI's own table behind it. Strategically, launch-week availability on both Microsoft Foundry and Amazon Bedrock lowers single-cloud risk, and Bedrock commits to no end of life before 2027-09-08. The risks are cost, a young model with a single undated snapshot, and a Critical cyber rating that brings monitoring that can halt agent sessions. Standardize it as the escalation tier above GPT-5.6 Sol and gate access per workload.
“OpenAI retakes the capability crown on math, science and web builds, while Anthropic keeps the agent and chat-preference edge for now.”
Astra leads where the benchmarks are hardest to game: Epoch's composite index, FrontierMath and ARC-AGI-3. Its WebDev first place turns that into a visible product edge. Claude Fable 5.1 still holds Arena's Agent top spot and a higher chat-preference score at the same list price. Gemini 3.1 Pro, still in preview at $2/$12, and Grok 4.7 at $2/$6 compete on price rather than peak capability. OpenAI's vertical launches (ChatGPT for Financial Services, Astra for Law) show where it will sell Astra: professional work. Timing is strong, since it arrived two days after Fable 5.1 and matched its price.
“At $10/$50 the list price matches Fable 5.1, but $1 cached reads and the 272K cliff make cache-heavy agents cost more than they look.”
Astra costs 2.5x GPT-5.6 Sol, 2x Claude Opus 5, 5x Gemini 3.1 Pro on input and more than 8x Grok 4.7 on output. At Fable 5.1's price, it loses on cached input ($1.00 against $0.25), which matters in long agent loops that re-read large prefixes. Cache writes bill at $12.50. Prompts over 272K double input and add 50% to output for the whole request. The offsets are real. Batch and Flex cut 50%, and Arena's median agent-task cost ($3.21) was lower than Fable 5.1's ($4.41) because Astra writes fewer tokens. Model the per-task cost on your own traces, not the rate card.
“Async tools, mid-turn steering and cache-safe effort switches are real upgrades, but plan for the Responses API and a model that asks back.”
The new controls fix real pain points in long agent runs. You can keep tools running while the model works, correct it mid-turn over WebSocket, and raise or lower effort without losing the cached prefix. The hosted tool set is complete, from computer use and hosted shell to apply_patch, MCP and tool search. Migration has sharp edges. none effort returns 400, sampling parameters and logprobs are gone, and function calling needs the Responses API. Astra also follows AGENTS.md and skill files more literally and stops to ask questions, so OpenAI publishes prompts to push it toward autonomy. There is no fine-tuning and only one snapshot.
“Excellent on hard problems and web builds, slower at high effort, and more inclined to ask a question or format everything as lists.”
On hard questions Astra is the strongest OpenAI model available, and Arena voters rank its web app output first. For everyday chat, preference data is still early. It sits at 1480 in Text Arena with a wide confidence interval, level with GPT-5.6 Sol. OpenAI itself says Astra tends toward detailed, list-heavy answers and recurring phrases. Its readiness to ask clarifying questions helps on ambiguous tasks and slows simple ones. Max effort answers take a long time. OpenAI is rolling it into paid and business ChatGPT products, but plan-level access varies, so check your plan.
“The 99.9% ARC-AGI-3 headline needs OpenAI's own harness. On the neutral one it is 62.7%, and human-preference chat data shows no lead yet.”
Independent data exists but is thin. Arena's text sample is small (2,693 votes, rank spread 5–51), and its coding sample is smaller still. ARC Prize reports 99.9% only with OpenAI's adapter that preserves hidden reasoning state. The neutral harness gives 62.7%. Astra's Epoch index lead (166.3 against Fable 5.1's 164.5) sits inside overlapping confidence intervals. xAI's own launch chart places Astra behind Grok 4.7 and Fable 5.1 on GDPval, though that is a rival's measurement. OpenAI's guide concedes pausing, heavy formatting and over-testing on small changes. OpenAI's eval table is not verified here, and SWE-bench-class scores are still missing.
none effort, no temperature/top_p/logprobs, and no function calling on Chat Completions.The full research notes behind this review — verified against primary sources.
OpenAI does not disclose parameter counts, layer counts, attention type, or whether Astra is dense or mixture-of-experts, so those fields are null. What is public: a single reasoning model with text and image input and text-only output, a 1,050,000-token context window with a 922,000-token input cap, and 128,000 output tokens. Reasoning tokens are hidden and returned only as summaries. Long sessions are managed with the platform's compaction and persisted-reasoning features rather than any disclosed architectural change. The only snapshot is gpt-6-astra, so there is no dated ID to pin yet.
Reasoning (9.9) and math (10) are the clearest strengths in independent data. Epoch AI's own runs give Astra 95.8% on GPQA Diamond, 97.6% on FrontierMath Tier 4, 93.7% on FrontierMath Tiers 1–3 and 100% on OTIS Mock AIME, each the best or tied-best score Epoch has recorded. It is also the first model to solve any FrontierMath Erdős problem (2 of 68).
Coding (9.9) rests on Arena. Astra is first on the WebDev board at 1800, ahead of Claude Fable 5.1 at 1758, and posts 1543 in Text Arena's coding category on a small vote count. OpenAI claims state-of-the-art software engineering, but its launch eval table is not reproduced here.
Agentic work (9.9) is backed by ARC Prize's result on ARC-AGI-3, an interactive benchmark: 62.7% under the provider-neutral harness and 99.9% with OpenAI's reasoning-state adapter, both state of the art. On Arena's Agent board Astra ranks second behind Fable 5.1. Function calling (9.5) adds async tools, programmatic tool calling and the full hosted tool set: web search, file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP and tool search.
Instruction following (9.3) is stronger by OpenAI's account, but Astra is more sensitive to AGENTS.md and skill files and more likely to stop and ask questions. Long context (9.0) reflects the 922K input cap and compaction support, with no independent long-context score yet. Vision (8.8) and document work (8.7) cover image input for computer use and document review; OpenAI's launch customer Legora reviewed 41 documents with it. Creative writing (8.6) is capped by OpenAI's own note that Astra leans on heavy formatting and recurring phrases. Multilingual (9.0) is a family-level estimate with no Astra-specific data. Safety calibration (9.0) reflects OpenAI's "most aligned model" claim and monitoring, balanced against the extra pauses. Real-time data (7.0) comes through the web-search tool.
OpenAI's own launch eval table is not reproduced here. Every figure below comes from an independent evaluator's page. Where an evaluator ran several effort levels, the level is noted.
| Benchmark | Score | vs GPT-5.6 Sol | vs Top Competitor | Source |
|---|---|---|---|---|
| GPQA Diamond (Epoch run, max) | 95.8% | 93.5% | Gemini 3.1 Pro Preview 94.4%, Claude Opus 5 93.9% | Epoch AI |
| SimpleQA Verified (Epoch run, max) | 75.6% | 69.7% | Gemini 3.1 Pro Preview 73.5%, Claude Fable 5.1 70.8% | Epoch AI |
| FrontierMath Tier 4 v2 (Epoch run, high) | 97.6% | 82.9% | Claude Fable 5.1 87.8% | Epoch AI |
| FrontierMath Tiers 1–3 v2 (Epoch run, max) | 93.7% | 89.1% | Claude Fable 5.1 90.2% | Epoch AI |
| OTIS Mock AIME 2024–2025 (Epoch run, max) | 100% | 100% | Claude Fable 5.1 100% (tie) | Epoch AI |
| FrontierMath Erdős (Epoch run, max) | 2 of 68 solved (2.9%) | 0 | Claude Fable 5.1 0 | Epoch AI |
| Epoch Capabilities Index | 166.3 (CI 163.0–171.9), #1 | 161.8 | Claude Fable 5.1 164.5 | Epoch AI |
| ARC-AGI-3 Semi-Private, Standard harness (max) | 62.7% ($26K run) | n/a | state of the art per ARC Prize | ARC Prize |
| ARC-AGI-3 Semi-Private, Provider Adapter harness (high) | 99.9% ($19K run) | n/a | state of the art per ARC Prize | ARC Prize |
| Arena Text, overall (max) | 1480 ±12, rank 24 (spread 5–51), 2,693 votes | 1483 (xhigh) | Claude Fable 5.1 1498, Claude Opus 5 1493 | Arena |
| Arena Text, coding (max) | 1543 ±23, rank 6 (spread 1–42), 645 votes | 1528 (xhigh) | Claude Opus 5 1533 | Arena |
| Arena WebDev (max) | 1800 ±16, rank 1 | 1617 (xhigh, Codex harness) | Claude Fable 5.1 1758 | Arena |
| Arena Agent, net improvement (max) | 11.54% ±2.10, rank 2 | 7.10% (xhigh) | Claude Fable 5.1 13.71% (rank 1) | Arena |
Arena snapshots are dated 2026-09-11 (WebDev), 2026-09-13 (Text) and 2026-09-15 (Agent). Epoch's evaluation was published 2026-09-03 after OpenAI gave Epoch pre-release access. MMLU-Pro, SWE-bench Verified, LiveCodeBench, Aider Polyglot, MMMU, Tau-bench, Terminal-Bench, HLE and the Artificial Analysis Index have no independently readable Astra score yet and are recorded null.
No independent tokens-per-second or time-to-first-token figure is recorded yet, so both fields are null. Latency depends mostly on the effort setting. low suits interactive use, while xhigh and max are built for long asynchronous runs. Fast mode doubles the token price for quicker processing, but OpenAI gives no latency SLA for Astra and does not offer Fast mode with EU data residency. Two independent signals point to a model that uses fewer tokens per task. On Arena's Agent board, Astra's median task used 27.9K output tokens against 49.2K for Claude Fable 5.1. ARC Prize found that higher reasoning levels often cost less in total because Astra solved games in fewer actions. latency_tier is slow because the effort levels that produce the headline results are slow.
| Surface | Cost | Notes |
|---|---|---|
| API input | $10.00 / 1M tok | prompts up to 272K input tokens |
| API output | $50.00 / 1M tok | hidden reasoning tokens bill as output |
| Cached input | $1.00 / 1M tok | 90% off input |
| Cache writes | $12.50 / 1M tok | 1.25x uncached input; set prompt_cache_options.ttl to "30m" |
| Long context (> 272K input) | $20.00 in / $2.00 cached / $25.00 cache write / $75.00 out | applies to the whole request |
| Batch (in/out) | $5.00 / $25.00 | 50% off |
| Flex (in/out) | $5.00 / $25.00 | 50% off, variable latency |
| Fast mode (in/out) | $20.00 / $100.00 | 2x; no latency SLA; not available with EU data residency |
| Regional processing | +10% | data-residency endpoints |
| Microsoft Foundry | $10.00 / $50.00 Standard Global; $11.00 / $55.00 Data Zone US | Provisioned Throughput also available |
| Amazon Bedrock | $10.00 / $50.00 Global CRIS; $11.00 / $55.00 in-region | Priority and Flex not supported |
| Direct UI | ChatGPT paid plans | plan-level access set by OpenAI |
| Free tier | none | no free API access |
| Rate limits | 15,000 RPM / 40M TPM | Tier 5; Tier 1 is 500 RPM / 500K TPM |
Reasoning-token note: at xhigh and max, hidden reasoning can bill several times the visible answer length. OpenAI says Astra needs substantially fewer output tokens than earlier models, giving a lower estimated cost per task despite the higher rates. That is a vendor claim, but Arena's Agent board supports it: median cost per task was $3.21 for Astra against $4.41 for Claude Fable 5.1.
Astra is served through the Responses API (https://api.openai.com/v1/responses), Chat Completions (without function calling) and Batch. It has no open weights, no self-hosting and no fine-tuning. Microsoft made it generally available in Foundry on 2026-09-03 with Standard and Provisioned Throughput deployments in Global and US Data Zone geographies. Amazon Bedrock added it on 2026-09-08 as openai.gpt-6-astra, with US and global cross-Region inference and a stated end of life no sooner than 2027-09-08. OpenRouter lists it as openai/gpt-6-astra. OpenAI's regional processing endpoints add 10% to the token price, and Fast mode is not available with EU data residency.
Astra is governed by OpenAI's Preparedness Framework, and OpenAI's safety overview calls it the first model to reach the Critical level of cybersecurity capability. Its release comes with strengthened safeguards and asynchronous misalignment monitoring on tool-using Responses API requests. These checks can raise safety alerts or stop a conversation for review, which production agent builders need to plan for. OpenAI does not train on API inputs by default, and opt-out and zero-retention options exist. The OpenAI platform covers SOC 2, GDPR and CCPA, with a HIPAA BAA available. On refusals, OpenAI describes Astra as its most aligned model. In practice the more visible change is that it asks clarifying questions more often and follows instructions in skills and AGENTS.md files more literally. Audit those files before migrating.
OpenAI's first-party SDKs cover Python, TypeScript, Java, Go and .NET, and the reasoning guide shows Astra examples in each. The OpenAI Agents SDK and common frameworks (LangChain, LlamaIndex, Vercel AI SDK) reach it through the standard OpenAI API. OpenAI publishes a Codex skill that applies its Astra migration steps to a codebase. Launch-period customer stories on OpenAI's site include Cognition's Devin, Perplexity, Hex, Legora, Playco and Higgsfield AI. OpenAI also built vertical products on it: ChatGPT for Financial Services and Astra for Law. Microsoft offers it in Foundry with its Agent Service, AWS serves it on Bedrock, and OpenRouter lists it. Popularity tier: mainstream, three weeks after launch.
$10 per 1M input tokens, $1 cached, $12.50 for cache writes and $50 output, with Batch and Flex at half price. That matches Claude Fable 5.1's list price, is 2x Claude Opus 5 and 2.5x GPT-5.6 Sol's current promotional rate.
The whole request bills at 2x input and cache rates and 1.5x output: $20 in, $2 cached and $75 out. Max input is 922K tokens even though the window is 1.05M.
Yes for plain text generation, but function calling with Astra requires the Responses API. OpenAI recommends Responses for all Astra work.
temperature, top_p, top_logprobs and logprobs are not supported, and none reasoning effort returns HTTP 400. Start at low if you ran GPT-5.6 at none or minimal.
Yes. Microsoft Foundry made it generally available on 2026-09-03 in Global and US Data Zone deployments. Amazon Bedrock added it on 2026-09-08 with in-region pricing 10% above OpenAI's rates.
Yes. On tool-using Responses API requests, OpenAI's asynchronous checks can trigger safety alerts or stop a conversation for review. Build handling for a halted run into production agents.
Both are far cheaper: Gemini 3.1 Pro Preview at $2/$12 and Grok 4.7 at $2/$6 with a 500K window. Astra beats Gemini 3.1 Pro on Epoch's GPQA Diamond (95.8% vs 94.4%) and index (166.3 vs 155.0). Grok 4.7 launched on 2026-09-21 with only xAI's own numbers so far.
API data is not used for training by default, and zero-retention options exist. Fine-tuning is not supported.
the same-price peer at $10/$50, released two days earlier. Fable 5.1 leads Arena's Agent board and Text Arena preference, and its cached reads cost a quarter of Astra's. Astra wins on Epoch's math and reasoning runs, Arena WebDev and ARC-AGI-3, and uses fewer tokens per agent task.
half the price at $5/$25 and the stronger value pick. Its Text Arena preference (1493) is higher than Astra's, but it trails Astra on GPQA Diamond (93.9% vs 95.8%), FrontierMath and WebDev (1687 vs 1800). It is the right default for teams that escalate to Astra only on the hardest work.
OpenAI's cheaper tier at $4/$20 (promotional through at least 2026-11-21), with the same 1.05M window and Pro mode. It is close on chat preference (1483 vs 1480) and much cheaper, but clearly behind Astra on Epoch's index (161.8 vs 166.3), FrontierMath Tier 4 (82.9% vs 97.6%) and Arena's Agent board.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-09-21