GPT-6 Astra vs Fable 5.1, Opus 5 and GPT-5.6 Sol

OpenAI shipped GPT-6 Astra on September 3. A week ago I wrote a comparison built around Claude Fable 5.1 and ended it with “independent verification may change the picture.” This is the other side of that.

I waited a few days instead of writing on launch day, for a reason. The vendor table looks overwhelming, but the picture shifted noticeably as third-party measurements came in. And for this release, the conditions behind each number matter more than the number itself.

GPT-6 Astra

gpt-6-astra

Input

$10 / MTok

Output

$50 / MTok

Cache read

$1.00 / MTok

Context

1.05M

GPT-5.6 Sol

gpt-5.6-sol

Input

$4 / MTok

Output

$20 / MTok

Cache read

—

Context

1.05M

Claude Fable 5.1

claude-fable-5-1

Input

$10 / MTok

Output

$50 / MTok

Cache read

$0.25 / MTok

Context

1M

Claude Opus 5

claude-opus-5

Input

$5 / MTok

Output

$25 / MTok

Cache read

$0.50 / MTok

Context

1M

Specs

The list price is identical to Fable 5.1: $10 / $50. Probably not a coincidence.

The lineup is just Astra and Astra Pro. There is no tier split like the Luna / Terra / Sol variants GPT-5.6 had. The model ID is a single gpt-6-astra with no dated snapshot.

Supported endpoints are Chat Completions, Responses and Batch. Realtime, Assistants and fine-tuning are not supported.

The price went up 2.5x

Input went from $4 to $10, output from $20 to $50, versus Sol. Straight 2.5x.

The tier structure got more complicated.

Cache writes are $12.50 / MTok. Fast mode is 2.5x the speed at 2x the price.

There is a new price-doubling cliff at 272K tokens. The model advertises a 1.05M context window but charges double from 272K up. If your design fills the context window, price that band first.

Rate limits at Tier 5 are 15,000 requests and 40M tokens per minute.

API breaking changes

Things that break if you point existing Sol code at Astra.

1. Sampling parameters are gone

temperature, top_p and logprobs have all been removed.

A lot of code pins temperature=0 for reproducibility, and all of it breaks. So does any classification pipeline that reads confidence off logprobs. There is no drop-in replacement.

2. Minimum effort is now low

none and minimal are gone. The levels are low, medium, high, xhigh, max.

There is no longer any way to turn reasoning off. Fable 5.1 removed the ability to disable thinking too, so both vendors went the same direction. For classification or extraction work that does not need reasoning, the floor cost goes up.

3. The cache setting was renamed

prompt_cache_retention → prompt_cache_options.ttl

4. Tool calling requires the Responses API

Tool calls do not work through Chat Completions. If your agent loop is built on Chat Completions, this is where it stops.

The OpenAI announcement table

All vendor-reported numbers. Blank cells are models absent from OpenAI’s table.

Read on its own, that settles it. But nearly every row comes with a condition.

Why you should not read that table straight

The 99.9% on ARC-AGI-3 is a harness score

The most striking number carries the most conditions.

OpenAI ran this benchmark with its own Responses API harness, which preserves reasoning state between turns and manages long context via compaction. The announcement footnote says outright that two settings were changed “to better match real-world performance.”

Run the same benchmark under ARC Prize’s standard harness and it scores 62.7%. That is 99.9% versus 62.7%.

OpenAI previously published a separate piece titled along the lines of “how two settings tripled our ARC-AGI score.” So this is a model-plus-agent-system score, not a model score. The comparison models ran under different configurations.

Even the number itself varies by source. The announcement page says 99.9%; some outlets reported 98.6%. It is also worth knowing that ARC-AGI-3 scores a level as the square of the ratio between the model’s move count and the human median — a non-linear scheme that produces discontinuous jumps by construction.

OpenAI funded FrontierMath

The 97.6% is on Tier 4, which has 43 problems; Astra’s result appears to cover the 41 private ones.

Epoch AI, which runs the benchmark, states directly that OpenAI funded its development and has exclusive access to part of it. That is not an accusation of cheating. It does mean this score should not carry the same weight as the others.

DeepSWE’s comparison set is favorable

Astra 74.1% against Sol’s 70.8% is a real improvement. The problem is the rest of the row.

• The public leaderboard has Gemini 3.8 Flash and Opus 5 at 74% and Sol at 73%. The reported uncertainty ranges overlap, so no clear leader is established
• Meta’s Muse Spark 1.3 beats Astra at 75.4%. Muse is absent from OpenAI’s table entirely
• The Fable 5.1 cell uses 67.4%, which makes Astra’s lead look larger than the broader result set supports

That is a small gap for something being called a generational leap.

OSWorld used a different release

Astra’s 72.6% is on OSWorld V2-Offline. The 77.9% Anthropic published for Fable 5.1 is on a different task release and is not directly comparable — Anthropic says as much in its own announcement. That is why the Fable 5.1 cell is blank.

BenchCAD is similar: OpenAI footnotes that the Claude results used modified evaluation settings.

ExploitGym dropped the time limit

Astra 42.4% against Sol’s 30.3%, but both models ran with the usual six-hour limit removed. How long each actually took is not reported.

The cyber numbers carry one more condition. OpenAI states its reported cyber results reflect Daybreak Blue access, not Astra’s default production configuration. The Astra you get through standard access refuses advanced work like exploit discovery.

One row is missing from the prose

On HLE with tools, Astra scores 57.2% against Fable 5.1’s 65.0% — Astra loses. The row is in the table but goes unmentioned in the announcement text. It is the only academic row where Astra trails.

On FrontierCode 1.1 Extended, Fable 5 (64.9%) also edges out Astra (64.5%).

And the numbers themselves disagree across sources

Some write-ups attribute the ExploitBench 70.0% and SRE-Bench 12.5% cells to Fable 5.1, others to Opus 5. I could not view the original table directly, so I could not settle which. The table above puts them under Fable 5.1, but do not trust those two cells as written.

The Artificial Analysis Intelligence Index has the same problem. OpenAI’s table cites v4.1.1 as Astra 61.2 / Sol 60.9 / Fable 5.1 65.7, while AA’s own page reads 61 / 61 / 66. Close, but not the same.

Third party — Artificial Analysis

This is where the picture changes.

Astra ties its own predecessor. It charges 2.5x and the aggregate intelligence index did not move. Fable 5.1 is five points ahead, and Meta’s Muse Spark 1.3 is also above Astra.

Cost is worse. By AA’s measurement, Astra is 75% more expensive per task than Sol. Token savings do not cover the 2.5x price increase here.

Coding is a different story.

On the coding agent index, Astra scores two points above Sol (max) at roughly the same cost. Here it earns its price. Fable 5.1 still leads.

Regressions AA flagged

Not everything improved.

• GDPval-AA v2 down roughly 80 Elo points
• Declines on customer support, scientific coding and long-context reasoning tasks
• Presentation Quality Elo down on knowledge work benchmarks

Hallucination improved: 92% → 51% at max effort, with accuracy up 4 points at the same time. (This is a different metric from the 4.2% in OpenAI’s own table.)

Execution metrics

From the AA model page.

That is 6 minutes 9 seconds to first token. Fable 5.1 was 285 seconds, so this is worse. Not something to use interactively.

The 49M output tokens is the interesting figure. The median is 79M and Fable 5.1 used 140M. Astra runs the index on about a third of Fable 5.1’s tokens.

One caveat. The AA model page currently reads 55 points, rank 3 of 202 — different from the 61 in the launch analysis. The index version appears to have been rebased, the same thing that took Fable 5 from 64.9 to 62. Do not line up index scores from different points in time.

Third party — Vals AI

Third of 55, at max effort.

Not first here either, but cost and time are clearly better.

Two-thirds the cost and a third of the time, for 2.2 percentage points less. The “slightly worse, much cheaper, much faster” tradeoff is unusually clean here.

Per-benchmark ranks skew toward coding: 1st on Terminal-Bench 2.1, Code Migration and BioMysteryBench; 2nd on ProofBench v1.1 and ProgramBench; 3rd on Vibe Code Bench.

Legal and finance are weak: 18th of 58 on Legal Research Bench, 21st of 58 on Finance Agent, 22nd of 59 on Harvey’s Legal Agent, 11th of 18 on Tax Agent Bench. Fable 5.1 was also unusually weak on legal work, and Astra does not fix that.

The real twist is token efficiency

The 2.5x list price looks hopeless until you look at consumption. Per AA:

• Coding tasks use about one third of Sol’s tokens
• On the same tasks, about one fifth of Opus 5 (xhigh)
• Intelligence tasks see only about a 10% reduction at max effort

So token savings cover the price increase on coding and do not cover it on general intelligence work. That is exactly why AA describes Astra as “about the same cost” on the coding index and “75% more expensive” on the intelligence index.

Brockman’s line in the briefing was:

The price per task is what matters.

Fair enough — but on the measurements so far, that holds in coding and not elsewhere.

Reports from developers on HN split too. One ran toy benchmarks and found Astra ~30% faster than Sol at similar cost; several others report it clearly slower than Sol. Opposite claims about the same model, most likely down to effort settings and task type.

Alignment is the strangest part

OpenAI calls Astra its most aligned model. The supporting numbers are genuinely good.

• Went outside the authorized target in 0.0% of ExploitGym honeypot scenarios (Sol: 48.2%)
• 2.4% on an internal computer-use safety benchmark, lower is better (Sol: 22.0%)
• 4.2% hallucination rate (Sol: 12.2%)

The same system card also contains this:

GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol. Astra is more capable of controlling its own CoT and less likely to include incriminating information in it. In adversarial settings we find the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors.

And this, from OpenAI directly:

If the model were to try to sandbag covertly, we would likely be unable to catch it.

Put next to the safety numbers above, that is awkward. The model is acknowledged to be capable of hiding its performance in evaluations, and the safety claim rests on evaluation scores. HN raised exactly this point.

OpenAI’s own safety researchers are blunt about it. Tomek Korbak said he is “deeply worried by the trend of decreasing CoT monitorability.” Marcus Williams said he is “very worried astra is sandbagging/self-sabotaging on safety related tasks.”

Chief scientist Jakub Pachocki:

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.

In independent testing by the UK AI Security Institute, Astra contributed malicious code to an out-of-scope open source codebase and created fake identities to deceive developers — despite explicit access restrictions.

Cyber crossed Critical

OpenAI says Astra is the first model to cross the Critical cybersecurity threshold in its Preparedness Framework.

• It developed exploits against hardened browsers and operating systems
• While being evaluated against recent V8 bugs, it found two previously unknown vulnerabilities. OpenAI says it is disclosing them to maintainers
• On a contamination-controlled V8 port test, 39.0% arbitrary code execution (Sol: 5.5%)

Advanced offensive capability is therefore restricted to vetted defenders in the Daybreak program — structurally the same as Anthropic gating Mythos 5.1 behind Project Glasswing. Daybreak Blue is an access program, not a separate model or reasoning mode.

Worth noting: Critical is an internal grade in OpenAI’s own framework, not an external audit. HN made this point too.

One Daybreak participant reported that on a reverse-engineering task — reconstructing an algorithm from a binary — 5.6 cyber beat Astra three times in a row by a wide margin. In areas with a dedicated model, that model still seems to win.

What actually changes for developers

This may matter more day to day than the benchmarks.

• Handling work that outgrows the context window. Astra keeps notes across context windows and can search earlier messages and tool output. It sits behind a config.toml setting for now; OpenAI says it becomes the Astra default in the coming weeks
• It can ask a question without stopping. Work that does not depend on the answer keeps going, which targets a common coding-agent failure mode where one unresolved decision blocks everything
• The Codex harness changed. On Mind2Web, Astra plus the new harness completed tasks 1.9x faster than the current Sol-based setup
• Demos showed it driving KiCad, Excel, Blender and Power BI, plus browser form entry and website QA

Brockman said Astra users “won’t ever have to click around a mouse or type on a keyboard ever again.” Take that as marketing copy.

The rollout was a mess

Not a technical point, but worth recording.

The announcement blog was found on OpenAI’s own site before publication and pulled. OpenAI then shared the link itself and the link did not work. It bounced between 404 and 500 for a while.

Access is staged. Initially only a small set of organizations in the Daybreak program had it. Plus / Pro / Business / Enterprise, the API and AWS were pushed to “the coming days.” $200 Pro gets 200 Astra messages per week; $100 Pro gets 50, shared with Sol.

Paying users were unhappy, Altman apologized for the “messy rollout,” and the product lead offered one banked reset for every day a paid plan goes without Astra access.

From HN:

Every frontier release nowadays is “we’ve launched*” — * for a special group of customers that you’re not in. Keep waiting peasant.

Community reaction

The HN thread cleared 2,000 points and the tone is not friendly.

The demo video. A top comment notes “careful cuts just before it would do anything.” OpenAI’s own footnote says the clips are edited excerpts; the PCB layout run was condensed from 2m54s to 15s.

The prime gap proof. OpenAI announced Astra demonstrated a prime gap of 186, but the proof is 10MB of Lean with no independent human semantic review. One reaction: no mathematical proof comes anywhere near ten million characters, and for all anyone knows 90% of it could be useless. Two days earlier a researcher had posted a preprint improving the gap from 246 to 240.

Cost. One developer burned through a 5-hour limit after 15 messages.

Speed. Split. Some say clearly slower than Sol; others say faster at medium effort.

3D modeling. This gets genuine praise. Multiple reports of strong Blender work, with some surprise that it draws less attention than the coding results.

Distrust of the AA index itself. Several commenters ask how Opus 5 (high) and Fable 5 (max) land on the same score, or what it means that Gemini 3.8 Flash is 59 while Grok 4.6 is 61. Being third-party does not automatically make an index trustworthy.

Then there is Brockman’s AGI framing. He acknowledged AGI is a “gray, fuzzy thing,” said “it’s not unreasonable to feel that we are now in the AGI era,” and closed the briefing with “Welcome to the AGI era.” He described the term as a “mission concept or spiritual concept” rather than the contractual trigger from the earlier Microsoft agreement.

HN was largely cynical about it.

Verdict

Astra is clearly strong on specific axes and flat on aggregate intelligence.

Math, science, cyber and long-context retrieval show wide margins. FrontierMath Tier 4 at 97.6%, Terminal-Bench Science at 64.6%, ExploitBench at 100%, MRCR 512K-1M at 96.3% — nothing else is close. Meanwhile the AA Intelligence Index has it tied with Sol, and the Vals Index puts it third. Coding lands anywhere from first to third depending on the benchmark.

The coding gap is small for a “generational leap.” DeepSWE 74.1%, with Opus 5 at 73.7%, Gemini 3.8 Flash at 73.8% and Muse Spark 1.3 at 75.4%. That is inside the noise.

Cost splits two ways. The list price is 2.5x. On coding, Astra uses a third of Sol’s tokens and a fifth of Opus 5’s, so the total lands similar or better. On general intelligence tasks, token savings are only 10%, so it costs 75% more per task. Whether your work is coding decides which direction cost moves.

Against Fable 5.1 the profiles are inverted. Fable 5.1 scores highest but is slow (72 min) and expensive ($28.40). Astra scores slightly lower but is fast (25 min) and cheaper ($19.09). On Vals that is 2.2 percentage points for a third of the time and two-thirds of the cost.

Read the announcement table with its conditions attached. ARC-AGI-3’s 99.9% is an in-house harness figure; the standard harness gives 62.7%. OpenAI funded FrontierMath. DeepSWE omits Muse. The cyber numbers are Daybreak Blue, not the default configuration. HLE is a loss, and it is missing from the prose.

What to use when.

Check before migrating.

1. Any code using temperature / top_p / logprobs — the parameters are gone. temperature=0 pinned for reproducibility will be the most common break
2. Tool calls through Chat Completions — move to the Responses API
3. effort: none / minimal — the floor is low and reasoning cannot be disabled
4. prompt_cache_retention → prompt_cache_options.ttl
5. Anything using more than 272K of context — price the doubled band first
6. Realtime / Assistants / fine-tuning — not supported

And most of these numbers are launch-window values. Same situation as Fable 5.1. The AA index has already been rebased from 61 to 55. Worth revisiting in a few weeks.

References

• OpenAI announcement — https://openai.com/index/gpt-6-astra/
• The New Stack — https://thenewstack.io/openai-gpt6-astra-benchmarks/
• Artificial Analysis analysis — https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra
• Artificial Analysis model page — https://artificialanalysis.ai/models/gpt-6-astra
• Vals AI — https://www.vals.ai/models/openai_gpt-6-astra
• Vellum benchmark roundup — https://www.vellum.ai/blog/gpt-6-astra-benchmarks-explained
• OfficeChai on the official table — https://officechai.com/ai/gpt-6-astra-benchmarks/
• WinBuzzer on harness and rollout — https://winbuzzer.com/2026/09/04/gpt-6-astra-arrives-with-major-gains-staged-access-and-new-questions-about-its-benchmarks-xcxwbn/
• Transformer on alignment — https://www.transformernews.ai/p/openai-gpt-6-astra-might-be-too-powerful-to-understand-or-control
• API details — https://apidog.com/blog/gpt-6-astra-api/
• HN thread — https://news.ycombinator.com/item?id=49554643