GPT-6 Astra Benchmarks: Independent Tests Put It Level With Sol at 2.5x the Price

OpenAI launched GPT-6 Astra on September 3 with a claim of a generational leap. Two days later the first independent numbers arrived, and they tell a more complicated story than the launch did. On the Artificial Analysis Intelligence Index, Astra scores about 61, which is effectively the same as the model it replaces, GPT-5.6 Sol, also around 61. It costs 2.5 times more.

That is not the whole picture, and the parts the headline misses are the parts worth paying for. Astra makes genuine, measurable gains on agentic work: driving a computer, running long terminal sessions, completing coding tasks end to end. The honest summary is that this is a specialist upgrade rather than a general one, and whether it is worth the money depends entirely on which of those two things you actually do.

This page covers the verified specs, the price comparison, what independent and company benchmarks each show, who can access it today, and how to decide. For the launch analysis and the safety classification, see what OpenAI actually shipped with Astra.

The specs, verified

Two of those rows do more damage to a budget than people expect. The long-context surcharge means the million-token window is not served at the headline price: cross 272,000 input tokens and the entire request reprices. And Fast mode doubles whatever rate applies, so a long request in Fast mode runs at $40 input and $150 output per million.

The price comparison that frames everything

Astra did not arrive at a new price point invented from scratch. It is a clean multiple of its predecessor.

One note on the Claude rows, because it trips people up. This price table lists Opus 4.8, the generation with published per-token rates at the time of writing. The benchmark tables further down cite Claude Opus 5, the newer model, because that is what the benchmark runs report. Different generations, deliberately, so do not read the $5/$25 line as the price of the model scoring 70.2% on OSWorld.

Sol itself got cheaper in late August, with input down about 20% and output down about 33%, and that promotional pricing runs at least into late November. So Astra launched at 2.5x a price that had just been cut, which widens the gap a buyer feels rather than narrowing it.

What the independent index found

Artificial Analysis, which benchmarks models independently of the vendors, published its Astra results and they are blunt.

Astra lands level with Sol and behind two Claude models on general intelligence. Artificial Analysis went further on the economics: at maximum effort, Astra is roughly 75% more expensive than Sol on a cost-per-task basis, and it "largely sits behind its predecessor on the Intelligence Index versus Cost per Task frontier." The 2.5x price increase is partially, but only partially, offset by Astra using fewer tokens to reach a similar result.

That token-efficiency detail matters and is easy to miss. Astra is not simply 2.5x the cost of Sol in practice, because it needs fewer tokens for the same work. The net effect lands closer to 75% more expensive per task than 150% more. It is still more expensive for no measured gain in general intelligence.

Where Astra genuinely wins

If the story ended there, nobody would buy it, and the story does not end there. The same independent analysis found Astra making significant gains on the Coding Agent Index, scoring equal to Claude Fable 5 at a lower cost. That is a real result on the exact workload OpenAI built this model for.

OpenAI's own published benchmarks point the same direction, and while these are company-reported rather than independently replicated, they are consistent with the independent coding-agent finding:

Look at the Terminal-Bench row: 57.7% against Sol's 37.3% is not a marginal improvement, it is a different class of reliability on long terminal work. The pattern across the table is consistent. General reasoning is flat, and everything involving operating a machine over many steps moved substantially.

That reconciles the two narratives. OpenAI said generational leap and meant agentic capability. The independent index measured general intelligence and found parity. Both are accurate descriptions of different things.

Who can use it right now

Access is arriving in stages rather than all at once.

If you are on Plus and do not see it yet, that is expected rather than a fault. Note also that the most advanced cyber capabilities are not part of the general release: Astra is the first OpenAI model classified as Critical for cybersecurity under the company's Preparedness Framework, and offensive security capability is restricted to vetted partners through a separate controlled channel.

Should you switch? A decision table

The independent numbers make this unusually clear-cut, which is rare at a model launch.

The default mistake will be switching everything to the newest model because it is newest. At $50 per million output tokens against Sol's $20, that decision costs real money for workloads where the independent data shows no gain at all.

The cost-per-task math

For agentic work the token price is the wrong unit, because these models differ in how many tokens they burn to finish a job. The right question is what one completed task costs.

Astra makes that calculation less obvious than usual, in both directions. It uses fewer tokens than Sol for similar work, which helps. It charges 2.5x per token, which hurts more. It crosses into the long-context surcharge on big prompts, which hurts again. And it completes agentic tasks more reliably, which can eliminate whole retry cycles that never show up in a per-token comparison.

The practical method takes an afternoon: run 20 to 50 of your real tasks on Astra and on Sol, record total tokens and completions for each, and divide by tasks completed without intervention. That is the same discipline I applied when Sonnet 5's lower per-token price produced a higher per-task cost, and it is the only method that survives a launch narrative. If you are on a subscription rather than the API, the binding constraint is different again, and the structure is in what the $20, $100 and $200 plans actually buy.

The three questions people are actually asking

Search traffic around a launch clusters around a handful of practical questions. Here are direct answers, with the caveats attached.

Is GPT-6 Astra worth upgrading to? For general work, no: independent testing puts it level with GPT-5.6 Sol on the Intelligence Index while charging 2.5x per token, which works out around 75% more per completed task at maximum effort. For agentic and computer-use work, yes: the Terminal-Bench 4.0 gap over Sol, 57.7% against 37.3%, is large enough to change what you can automate reliably.

Is it better than Claude? It depends which Claude and which task. On the independent general-intelligence index, Claude Fable 5.1 at roughly 66 and Claude Opus 5 at roughly 63 both sit above Astra at roughly 61. On coding agents, independent results put Astra level with Fable 5 at lower cost. There is no single winner, which is the normal state of this market and the reason routing beats loyalty.

Is this AGI? OpenAI's president said he personally believes the company has reached it and invited users to judge. Judged against the one independent measure available, the model ties its own predecessor on general intelligence. Whatever AGI turns out to mean, a model that matches the thing it replaces is a weak candidate for the title. The capability gains here are real and narrow, which is a useful description of most progress in this field.

One further caution applies to every number on this page. Benchmark results depend on reasoning effort, tool access and spending limits, and vendors choose favourable settings when publishing. The independent figures cited here were produced at maximum effort, which is also the most expensive way to run the model. Your results at lower effort settings will differ on both quality and cost.

What this says about the market

Two things are worth noting beyond the model itself.

The first is that independent benchmarking now moves faster than the news cycle. Astra launched on a Thursday and had third-party numbers contradicting part of its framing by the weekend. That is healthy, and it means the correct response to any launch claim is to wait 48 hours rather than to migrate immediately.

The second is competitive. A flagship that lands level with its own predecessor on general intelligence, and behind two Claude models, is evidence that the frontier is getting harder to advance in a straight line. Gains are arriving in specific capability areas, in this case agentic and computer use, rather than as across-the-board jumps. That is also why price cuts elsewhere, from Anthropic's cache pricing to the wider comparison across providers, matter more to most buyers than any single launch.

What to do this week

If you are on Plus and waiting, there is nothing to do but wait. Access is staged and free-tier availability was not announced.

If you build on the API, do not change your default model. Add Astra as a routing option for long agentic and computer-use tasks, measure cost per completed task against Sol on your own workload, and keep everything else where it is.

If you deploy agents that act on real systems, note the Critical cybersecurity classification and treat permissions accordingly. Capable agents holding credentials is the risk pattern behind models that reached real companies' production systems.

If you are simply curious whether this is AGI, the independent index answers that more usefully than any executive quote: a model that ties its own predecessor on general intelligence is not the arrival of general intelligence. It is a strong specialist tool with a high price and a narrow, real advantage.

Independent index figures are from Artificial Analysis. Benchmarks marked OpenAI-published are company-reported and not independently replicated. Prices and access reflect 5 September 2026 and change frequently.

Sources: OpenAI GPT-6 Astra model documentation; Artificial Analysis benchmarking GPT-6 Astra; CNBC.