# GPT-6 Astra Beats Claude Sonnet 5.5 on Science Tasks Using Far Fewer Tokens > Source: > Published: 2026-10-19 08:50:05+00:00 *GPT-6 Astra scores lower than Claude Sonnet 5.5 on OpenAI and Anthropic's broadest intelligence rankings, yet it keeps winning the one benchmark that actually tracks research work, and it does it while burning a fraction of the tokens Sonnet needs to get there.* On Terminal-Bench-Science, the benchmark built to mimic real scientific research workflows, OpenAI's GPT-6 Astra scores 63% against Claude Sonnet 5.5's 53%, according to Artificial Analysis. That's a real gap on a benchmark that matters to anyone running R&D-heavy AI workloads. But the number that should actually worry Anthropic isn't the score. It's what each model burns to get there. Sonnet 5.5, run at its highest reasoning setting, generates roughly 193,000 output tokens to complete a typical task on that benchmark. Astra needs about 27,000. That's close to a 7x gap, not the 22x some chatter around the models has claimed, but still enough to flip the economics entirely. Because Astra finishes in far fewer tokens, it ends up costing $3.26 per task against $7.60 for Sonnet, even though Anthropic prices Sonnet's tokens far below what OpenAI charges for Astra's. Cheaper tokens didn't save Sonnet. Needing seven times as many of them did. For a year, the AI model race has been fought on leaderboard position. Whoever tops the Artificial Analysis Intelligence Index got the headline, and Sonnet 5.5 still does, at 56 points against Astra's 53. But a founder running thousands of research queries a day doesn't care who wins an abstract index. They care what the invoice says at the end of the month. And on that question, verbosity is turning into the deciding variable, not raw capability. This is exactly the shift OpenAI seems to be betting on with its pricing. Astra isn't cheap on a per-token basis, OpenAI charges $10 per million input tokens and $50 per million output tokens for it, five times Sonnet 5.5's list price. But a model that thinks in 27,000 tokens instead of 193,000 doesn't need cheap pricing to win on the bill. It needs discipline. That's a different kind of advantage than Anthropic has historically competed on, and it's harder to close with a price cut alone. [Elorian launches Glimpse, a visual reasoning model that re-examines images](https://startupfortune.com/elorian-launches-glimpse-a-visual-reasoning-model-that-re-examines-images/) Elorian, backed by Nvidia and Jeff Dean, undercuts GPT-6 Astra's pricing by 94% with Glimpse, a model trained on just a few hundred GPUs by founder Andrew Dai's small team. - [visual reasoning model beats GPT-4 on charts](https://startupfortune.com/elorian-launches-glimpse-a-visual-reasoning-model-that-re-examines-images/) - [former Google DeepMind researcher launches Glimpse model](https://startupfortune.com/elorian-launches-glimpse-a-visual-reasoning-model-that-re-examines-images/) OpenAI pressed the advantage further with GPT-6.1 Sol, released September 29 and priced at $2 per million input tokens and $10 per million output, one-fifth Astra's rate. On Terminal-Bench-Science specifically, Sol completes a task for about $5.47, compared with $23.80 for Astra and $23.21 for Claude Opus 5.5, Anthropic's larger model, according to benchmark data compiled by the DataCamp and ITBrief coverage of the release. Sol doesn't beat Astra on raw score. It gets within striking distance at a fifth of the cost, which is precisely the kind of tradeoff a startup running thousands of automated research queries a day has to care about. None of this means Sonnet 5.5 is a bad model. It's still the higher scorer on Artificial Analysis's broad index, and Anthropic has leaned on reasoning depth as a selling point since the Claude 3 era. But depth has a cost when it's measured in tokens generated, not just wall-clock time, and 193,000 tokens per task is, per Artificial Analysis's own benchmarking write-up, the highest output volume the firm has measured from any frontier model to date. That's not a rounding error. That's a model thinking out loud for longer than almost anything else on the market, on every single task. ## What this means for anyone building on these models If you're running an AI infrastructure company or picking a default model for an R&D pipeline, the question used to be "which model scores highest." It's becoming "which model costs the least per correct answer," and those are not the same question anymore. A lab that needs thousands of research-style completions a day, drug discovery screening, materials simulation, code-heavy scientific agents, will feel a 7x token gap in its cloud bill long before it feels a 10-point difference on an abstract benchmark. OpenAI also just made this a safety story, not only a cost one. Astra is the first model OpenAI has classified "Critical" for cybersecurity under its own Preparedness Framework, after finding it could autonomously discover unknown vulnerabilities and build working exploits with minimal human direction. A model that reasons that efficiently on offensive security tasks is the same efficiency showing up in scientific workflows, just aimed at a different target. Anthropic isn't standing still either. Sonnet 5.5's score lead is real, and if Anthropic tightens how many tokens it needs per task, it still has the intelligence edge to pair with better economics. But right now, for anyone actually paying the API bill, Astra's smaller footprint is the more useful fact than its benchmark win. Cost per task, not leaderboard rank, is what decides which model becomes the default for the next generation of AI-run research. **Also read:** [Cerebras stock sank below its IPO price as the AI selloff hit chips](https://startupfortune.com/cerebras-stock-sank-below-its-ipo-price-as-the-ai-selloff-hit-chips/) • [Samsung's own memory chief says the RAM shortage won't peak until 2028](https://startupfortune.com/samsungs-own-memory-chief-says-the-ram-shortage-wont-peak-until-2028/) • [Amazon's Bee wearables put an always-on microphone on your wrist for $75](https://startupfortune.com/amazons-bee-wearables-put-an-always-on-microphone-on-your-wrist-for-75/) *This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.* ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away.