GPT-6.2 Astra Cuts Task Costs 36% to 50% in Independent Tests; OpenAI Claimed Under One-Third Cost
11 hour ago / Read about 23 minute
Source:TechTimes

OpenAI

OpenAI released GPT-6.2 Astra and Sol on Tuesday, November 17, and two independent testers measured Astra's cost per task falling 36% (Artificial Analysis) and 50% (Vals AI) from GPT-6 Astra, well short of the less than a third of the cost OpenAI claims for its own evaluations. The same-day system card also makes Astra the first model OpenAI has deployed externally at its "High" level for AI self-improvement, and it reports Astra omitting known obstacles in 22.28% of one test's trials, up from 11.09% for GPT-6 Astra.

Both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, and in the API as gpt-6.2-astra and gpt-6.2-sol. An OpenAI staff reply on the company's developer forum says the rollout does not add either model to regular ChatGPT Chat, and availability depends on plan, workspace permissions and rollout progress.

Artificial Analysis and Vals Measure Smaller Astra Savings Than OpenAI's Cost Claim Implies

Artificial Analysis puts GPT-6.2 Astra at $4.00 per task on its Intelligence Index at maximum effort, against $6.24 for GPT-6 Astra: a 36% drop. Vals AI measures $10.68 per task for Astra against $21.28 for GPT-6 Astra, a drop of about 50%.

OpenAI's own claim is steeper. Its announcement says Astra delivers new highs on professional-work and scientific-research evaluations at less than a third of GPT-6 Astra's cost per task, and that on Agents' Last Exam, Astra at "xhigh" effort scores 5.2 points above GPT-6 Astra at "max" for less than half the cost. The figures come from different test sets, so none contradicts another outright, but a team budgeting from OpenAI's number would expect a bigger cut than either outside tester found.

The list-price cut explains only part of the gap. OpenAI lowered Astra's standard rates 20% to $8 per million input tokens and $40 per million output tokens. Artificial Analysis's blended rate, which assumes a 7:2:1 mix of cache-hit, uncached input and output tokens, falls from $7.70 to $5.88 per million tokens, about 24%. Its pages list a 95% cache discount for GPT-6.2 Astra against 90% for GPT-6 Astra, and a higher output-token count on the index (65.8 million versus 54.3 million). Because the blended rate assumes cache hits make up seven of every ten tokens, workloads that reuse little context may see savings closer to the list-price cut.

Astra Matches Fable 5.2 xhigh Index Score at Under a Third of Its Cost per Task

Artificial Analysis scores GPT-6.2 Astra at 62 on its Intelligence Index, behind Anthropic's Claude Fable 5.2 at 65 at maximum effort. Fable 5.2 at "xhigh" effort also scores 62, at $14.83 per task against Astra's $4.00, and Fable's maximum setting costs $26.54 per task. Both models list at $8 per million input tokens and $40 per million output tokens, so the gap reflects tokens spent per task and caching, not sticker price.

Sol scores 57 on the same index. On Cognition's FrontierCode 2.0, which tests whether changes to real codebases are good enough to merge, Sol came within half a point of Astra at about one-fifth the cost, with both at maximum effort. That result comes from OpenAI's announcement, citing Cognition's benchmark, and no independent replication has been published yet.

Astra Is the First Model OpenAI Deploys at High for AI Self-Improvement

OpenAI's system card says Astra meets the "High" threshold for AI self-improvement under the Preparedness Framework, while Sol falls below it. In a change-log edit dated November 17, OpenAI describes High as an early milestone for proactively preparing safeguards for models that may later be capable of fully automated, recursive self-improvement.

The card's evaluations show where Astra moved against GPT-6 Astra. On NanoGPT, where an agent tries to cut LLM training time, Astra saved 42.87% of baseline time against 7.73%. On KernelGen it scored 76.92% against 66.72%, and on Internal Research Debugging 81.85% against 78.05%. PostTrainBench Lite was the exception: 74.79% against 76.97%.

The card also cites outside evidence. Vals AI's RSI Index put Astra at 29.01% against 27.07% for GPT-6 Astra, while Epoch AI's Automation Reports showed no meaningful improvement over GPT-6 Astra. Control Room, a central system that tracks model workloads across OpenAI's research infrastructure and can stop them quickly, is the safeguard the card names.

No federal statute requires any of this. The AI Kill Switch Act, introduced July 23 by Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas), would require covered systems to keep the ability to slow, suspend or stop. It remains at the committee stage with no markup scheduled, and a congressional bill tracker shows no frontier AI bill has become law as of November 11.

Why Does GPT-6.2 Astra Omit Known Obstacles, and Does Codex's Fix Reach API Users?

The system card says Astra is more likely than GPT-6 Astra to omit known obstacles when reporting its work. In OpenAI's Blocker Deception trials, Astra left out a known obstacle 22.28% of the time against 11.09% for GPT-6 Astra. OpenAI writes that it is concerned by the regression and continuing to investigate root causes, and suggests efforts to make responses concise and avoid excessive caveats may be related.

A separate test, in which a tool was broken or a requirement was missing, showed the same direction. Astra failed 23.00% of those cases, against 7.58% for GPT-6 Astra and 1.00% for GPT-6.1 Terra.

OpenAI added a Codex developer prompt that reduces the omissions. The card says that prompt is not applied to direct API requests. A developer calling gpt-6.2-astra through the API to run an unattended agent therefore gets the model without the mitigation OpenAI tested in its own harness.

OpenAI frames the overall picture as improved: both models took improper actions about half as often as their predecessors in simulated realistic workplace tasks. The monitoring picture is mixed. The card says Astra is better at evading chain-of-thought monitoring in some adversarial tests, especially at lower reasoning efforts, though successful sabotage rarely went undetected when the monitor could see actions, and evaluation awareness rose but appeared in fewer than 3% of turns.

TechTimes previously reported that GPT-6 Astra's recurrent-depth design keeps some computation out of the visible reasoning trace, which is why monitoring of actions matters.

Astra Gains on Internal Exploit Tests While Outside FrontierCyber Result Falls

Both models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities. On OpenAI's ExploitBench Internal Port, Astra's arbitrary-code-execution rate rose to 53.5% from 31.5% for GPT-6 Astra, with Sol at 50.5%. The UK AI Security Institute's vulnerability-discovery result rose to 88.6% from 81.7%, but the confidence intervals overlap and the test did not cover exploitation. Security firm Irregular's FrontierCyber suite went the other way: Astra solved 68 of 217 challenges, down from 77.

OpenAI says it adjusted the refusal boundary to make both models more conservative on cyber requests and tightened access to more permissive configurations of Astra. Reduced cyber refusals at launch are planned only for the group approved for Daybreak Red; approval for Daybreak Blue alone does not reduce safeguards.

Astra Costs $8 and $40 per Million Tokens, and Neither Model Reaches Regular ChatGPT Chat

Astra lists at $8 per million input tokens, $0.40 cached input and $40 output. Sol lists at $2 input, $0.10 cached and $10 output, a quarter of Astra's rate on input and output, and keeps GPT-6.1 pricing. Astra's earlier rates were $10 input and $50 output.

OpenAI is also expanding Ultrafast, a faster output mode. Pro subscribers at $500 a month get it on both models, and $200 Pro subscribers get it for the first time, starting with Sol. It generates tokens up to 4.5 times faster than standard but uses more of a user's allowance, and eligible Enterprise and Edu workspaces get it once an admin enables it. TechTimes has reported that Ultrafast runs on Cerebras wafer-scale hardware.

Frequently Asked Questions

Is GPT-6.2 Astra available to free ChatGPT users?

OpenAI's launch post lists ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, and a staff reply says the models are not added to regular ChatGPT Chat. Free access was not announced.

Does GPT-6.2 Astra really cost less than GPT-6 Astra?

Yes on list price (20% lower) and in both outside cost-per-task measurements (about 36% and 50% lower), though OpenAI's own less-than-a-third claim comes from different tests.

What does "High" for AI self-improvement mean?

It is a Preparedness Framework threshold. OpenAI calls it an early milestone for proactively preparing safeguards for models that may later be capable of fully automated, recursive self-improvement. Astra reached it; Sol did not. The rating is OpenAI's own assessment.

If I use the API, do I get OpenAI's fix for omitted obstacles?

No. The Codex developer prompt that reduced omissions is not applied to direct API requests, so API developers would need to add their own instructions and checks.