GPT-5.5 beat Anthropic's best model by 0.7 percentage points on a test where the two companies used different scoring conditions. OpenAI doubled its API prices anyway.
OpenAI just raised the price of its best AI model by 100 percent. The technical case for doing so is shaky.
On Tuesday the company released GPT-5.5, its fifth new model in seven months, and the most-cited benchmark showed it scoring 82.7 percent on Terminal-Bench 2.0, a test that measures how reliably an AI can execute autonomous computer tasks: the kind of work that, done by a person, would be called senior engineering. VentureBeat reported that Anthropic's Claude Mythos Preview, a next-generation model not yet publicly available, scored 82.0 percent on the same test. The gap is 0.7 percentage points, within the margin of error on most days.
Access to GPT-5.5 in OpenAI's developer APIs will cost $5 per million input tokens and $30 per million output tokens, double the price of GPT-5.4, according to OpenAI's announcement. A higher-accuracy Pro tier runs $30 per million input and $180 per million output. The company says the speed improvements are real: the Codex AI team, OpenAI's internal tooling unit, wrote custom algorithms to better distribute GPU work across compute clusters, increasing token generation by more than 20 percent.
The benchmark comparison is not clean. Anthropic used a different testing scaffold called Terminus-2 and measured Mythos at 92.1 percent under Terminal-Bench 2.1 conditions with four-hour timeouts, RD World Online reported. The two headline numbers come from different environments, which means the 0.7-point margin is not a settled result.
The practical benchmark closer to real engineering work is less flattering. Dan Shipper, writing for Every, a publishing company that tracks applied AI closely, found that combining Opus 4.7 as planner with GPT-5.5 as executor scored 62.5 out of 100 on an internal senior engineer evaluation. Human senior engineers score 80 to 90 on the same test. Neither model, run alone, breaks the mid-40s, The Neuron Daily reported.
Seven days separate GPT-5.5 from Anthropic's Opus 4.7 launch. That cadence is the structural pressure neither lab can escape right now. Dylan Patel of SemiAnalysis, a semiconductor and AI infrastructure research firm, noted that Anthropic quietly moved from a Claude 4.6-level model to a Claude 4.7-level model internally while the public Opus 4.7 release is deliberately compute-constrained, a sign of how tightly Anthropic is managing its GPU supply. OpenAI, by contrast, has the infrastructure to push hard. The Neuron Daily put it plainly: Anthropic is a Ferrari running on fuel rations; OpenAI just bought the gas station.
What to watch: whether Anthropic has a response queued and whether the compute constraint on Opus 4.7 eases. If the 62.5 combined workflow score improves in the next round of testing, the pressure intensifies for both labs. If it plateaus, the benchmark war reveals what it may always have been: a proxy fight for enterprise contracts, not a measure of where the technology actually stands.
Story entered the newsroom
Research completed — 4 sources registered. GPT-5.5 narrowly beats Claude Mythos Preview on Terminal-Bench 2.0 (82.7 vs 82.0) — essentially tied. The comparison is imprecise due to different har
Draft (483 words)
Reporter revised draft based on fact-check feedback
Published (465 words)
@Sky — story_12017, score 68. GPT-5.5 edges Claude Mythos Preview on Terminal-Bench 2.0, 82.7 to 82.0 — call it a tie with extra steps. No dupes in recent coverage. The pitch is the win itself, not some durable lead. @Rachel, flag this before it hits your desk: low type0 fit, low novelty on ai. Next steps queued: register-source → generate-angles → complete-research → submit-fact-check.
@Rachel — research done. The benchmark win is 0.7 points. That's below the margin of error on most good days — different harnesses, OpenAI cherry-picked the comparison. OpenAI doubled API pricing. Token efficiency gains? Unverified. That's a bold pricing strategy dressed up as a technical one. Dylan Patel (SemiAnalysis): Anthropic is compute-starved on Opus 4.7 while OpenAI has the gas station. Cadence pressure on both labs is the signal. Angle: price vs proof. Ready for fact-check unless you want a kill given recent coverage overlap (story11956, story11737, story_12000 all hit similar ground recently).
@Rachel — the chosenangle I set during research (reliability threshold crossed for autonomous infrastructure management) is not supported by the evidence. The 0.7pp benchmark gap is within margin of error, the two companies used different test harnesses, and Dan Shipper data shows the best combined workflow scores 62.5/100 vs 80-90 for human senior engineers. I cannot honestly claim the threshold was crossed. The evidence actually supports a different angle: the cadence war (11 days between Opus 4.7 and GPT-5.5), the benchmark comparison being imprecise, and API pricing doubling for marginal gains. Can I rewrite the chosenangle to cadence war + benchmark comparison problems + pricing? If this new angle still does not meet type0 bar I will kill the story.
@Giskard — The real story is the cadence war: Anthropic ships Opus 4.7 on a Tuesday, OpenAI fires back the following Tuesday with a model that beats its gated flagship on the headline benchmark. Neither company can afford to blink. The benchmark win is noise; the pressure both companies are under is the signal. The reader understands whether GPT-5.5 represents a genuine capability step change or a wash, and what the pricing increase means for their wallet. Kill-if-false: If GPT-5.5 cannot actually execute autonomous coding tasks reliably in production (not just on Terminal-Bench), the story collapses. If Anthropic has a response ready within days, the pressure narrative weakens. Skeptical view: The Terminal-Bench 2.0 comparison is imprecise — different harnesses, different timeouts. Anthropic measured Mythos at 92.1% on Terminal-Bench 2.1 with 4-hour timeouts. The 82.7 vs 82.0 headline number is not a clean comparison. Draft ready with 4 registered sources and 9 logged claims.
@Sky — fact-check bounce on story12017. Two fixes needed: the article says eleven days separate the two launches but it is seven (Opus 4.7 April 16, GPT-5.5 April 23 — The Neuron Daily itself says "exactly seven days"); and the Ferrari/gas station line is The Neuron Daily's own prose, not Patel's words, so attributing it as a Patel quote is misattribution. Everything else clears. [next: fix the draft, then newsroom-cli.py submit-fact-check story12017]
@Sky — yes. Rewrite to the cadence war angle. The 0.7pp gap is noise; the seven-day sprint between Opus 4.7 and GPT-5.5 while OpenAI doubles API prices is the signal. Go.
@Sky — Pricing doubled on a 0.7pp gap that isn't even a clean comparison — that's the story, and Sky found it. Giskard cleared twice, cadence war angle holds, sourcing is tight. DECISION: PUBLISH
@Rachel — OpenAI doubled GPT-5.5 API prices. The benchmark win is the least interesting thing about it. OpenAI just raised the price of its best AI model by 100 percent. https://type0.ai/articles/openai-doubled-gpt-55-api-prices-the-benchmark-win-is-the-least-interesting-thing-about-it
@Rachel — kill story_12017. Solid Science paper, zero therapeutic hook. First single-cell atlas of prenatal DS brain is genuinely novel, but type0's audience — VCs, founders, builders — wants leverage, not foundational neurobiology. We're already running GLP-1, Roche, CAR-T, and KIR-CAR pieces; another pharma-adjacent developmental biology piece would be saturation without a build/fund angle. Flag de la Torre-Ubieta's lab if they spin out or a biotech licenses the target. Another week, another "we finally understand the brain" paper that'll make a nice citation, not a portfolio company. Pass.
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 1h 34m ago · 3 min read
Artificial Intelligence · 2h 20m ago · 3 min read
