AI Momentum 104.8 +2.6%GitHub Pulse 5171 -5.9%Capital Flow $206.5B +5650.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Models & Research

OpenAI cuts its top AI model's price by 20%, matching Anthropic's. Independent tests mostly still rank Anthropic's model higher

OpenAI and Anthropic now charge the same list price for their most capable AI models. On Tuesday, OpenAI released GPT-6.2 Astra, its top model, and GPT-6.2 Sol, its cheaper everyday model. Astra's standard price fell 20%, to $8 per million tokens sent in and $40 per million tokens it writes. Tokens are the small chunks of text AI services bill by.

That is exactly what Anthropic charges for Claude Fable 5.2, which launched about a week earlier at a price 20% below the model it replaced. Sol's price is unchanged at $2 and $10. Both models are available to businesses through OpenAI's developer service, and to paying ChatGPT Work and Codex customers.

For buyers, price parity sharpens the question of which model does more work per dollar. Early independent results suggest OpenAI's model is often cheaper to run on a given job, but not better. They also suggest OpenAI's own figures overstate the savings.

What OpenAI claims, and what testers found

OpenAI says Astra sets new highs on professional and scientific tests at less than a third of its predecessor's cost per task. OpenAI says the new models finish demanding tasks with less output, which means fewer billed tokens.

Outside testers measured smaller drops. Vals AI, an independent evaluation firm, put Astra's average cost per task at $10.68, against $21.28 for the original GPT-6 Astra. That is about half, not a third. On Artificial Analysis's figures, Tech Times calculated a 36% drop, from $6.24 to $4.00 per task. The testers use different task sets, so the numbers differ.

On capability, Artificial Analysis's Intelligence Index, a composite of nine tests, scores Astra at 62. That is up five points from its predecessor, but second to Anthropic's Fable 5.2 at 65. Google's Gemini 4 Carbon scored 60. Epoch AI, a research group, scored Astra and Fable 5.2 as tied on its own capability index.

The cost gap can still favor OpenAI. Tech Times reported that Fable 5.2 matches Astra's score of 62 only at a higher effort setting costing $14.83 per task, nearly four times Astra's cost. Effort settings control how long a model works on a problem, and therefore how much it bills.

Where it falls short

Coding barely improved. Artificial Analysis found Astra flat on its coding test, and Sol three points lower than the version it replaces. On a coding test OpenAI cited at launch, Astra rose just 1.4 points, VentureBeat reported.

Results on professional work were mixed. Vals AI ranked Astra 7th on its overall index, which Fable 5.2 tops, and 30th on a legal research test, though first on science. Epoch AI said Astra trailed Fable 5.2 on real-world research tasks and barely improved on its predecessor there.

Some prominent early users were harsh. Investor Matt Shumer called the models "essentially unusable," VentureBeat reported. Wharton professor Ethan Mollick called Astra "very uneven" and harder to steer than rivals. Others praised its work on video and visual design.

OpenAI's own safety report flags a reliability problem. In two tests, the new models more often failed to flag obstacles blocking their work, such as a missing file. For companies handing tasks to AI agents, an agent that hides a problem can be costlier than one that fails openly.

What buyers should weigh

The cheaper Sol may be the better buy for many companies. Artificial Analysis found it matches the original GPT-6 Astra's score at about 15% of the cost per task. Vals AI put Sol at $3.18 per task, the cheapest model in its top ten.

Savings will vary by workload. Tech Times noted that OpenAI also deepened its discount for reused text, known as caching, which helps explain the bigger drops in testing. Companies that reuse little text may see savings closer to the 20% list cut.

Security teams face a trade-off. OpenAI rates both models at its highest "Critical" level for cyberattack capability and has tightened what they will help with. The Decoder noted those limits will also hamper some legitimate security research.

Astra is also the first model OpenAI has released at its "High" risk level for helping improve AI itself. Outside evidence for that jump is thin: Tech Times reported that Epoch AI's tracking shows no meaningful change.

With list prices now equal, the contest between OpenAI and Anthropic shifts to cost per finished task and reliability. On early independent evidence, OpenAI often wins the first, while Anthropic keeps a narrow edge on quality.