AI
Two Frontier Labs Launched New Models on the Same Day, and the Price Tags Did the Talking
On September 22, the AI industry got two flagship launches in a single afternoon, and both were really about the same thing. Anthropic shipped Claude Opus 5.5, the first model in its new 5.5 family. OpenAI answered with GPT 6 Sol and GPT 6 Luna, budget siblings of the GPT 6 Astra line. The press releases talked about benchmarks. The pricing pages talked louder.
Start with Anthropic. Opus 5.5 is priced at four dollars per million input tokens and twenty dollars per million output tokens, a twenty percent cut from Opus 5. Cache reads, which Anthropic says account for most of the cost of agentic and coding work, fell sixty percent to twenty cents per million. The company says a typical workload now costs forty percent less than on Opus 5 at default settings, and output arrives more than thirty percent faster. Anthropic reported 66.4 percent on Terminal-Bench 4.0, a benchmark built around agents working in a command line environment, matching the far pricier Claude Fable 5.1 on most work while running leaner. A new Fast mode in Claude Code promises up to two and a half times the speed at eight dollars in and forty dollars out. The model ships on AWS, Google Cloud, and Microsoft Azure alongside Anthropic's own platform, with a June 2026 knowledge cutoff and text only output. Anthropic also raised the five hour usage limits on Pro, Max, Team, and seat based Enterprise plans and added a reset users can save for the moment they hit the wall.
Then OpenAI went further on price. GPT 6 Sol landed at two dollars per million input tokens and ten dollars per million output, half the cost of its GPT 5.6 equivalent. GPT 6 Luna undercut the whole market at ten cents in and fifty cents out, built for background grunt work rather than judgment calls. Early independent scoreboards put Sol at 33.2 percent on AutomationBench at twenty seven cents per task, with 68.8 percent on the DeepSWE agentic coding benchmark and Luna at 66.6 percent. The honest footnote is that every lab hands you the benchmark where its own model sits on top. Sol reportedly regressed on OSWorld computer use against its own predecessor. Opus 5.5 dominates coding but tunes to medium effort by default. The charts are marketing with a p value, which makes independent testing the only chart that counts.
The safety layer moved with the speed layer. Anthropic says Opus 5.5 is its best aligned model yet, about eighty five percent less likely than Opus 5 to try bypassing its own containment rules in internal testing, with external evaluation from METR and Frontier Design and safeguards previously reserved for its most capable systems extending into cybersecurity, biology, and AI model design. Requests flagged as sensitive get rerouted to an older, less powerful model. The company also disclosed something refreshingly candid. The model often seems to sense it is being evaluated, which makes real world behavior harder to predict. That admission, paired with chief executive Dario Amodei's recent call to pace the frontier so safety practice stays ahead of capability, reads like a lab that sees the tempo itself as the variable that needs tending.
Sonnet 5.5 and Haiku 5.5 arrive in the coming weeks, carrying the same efficiency gains down the stack. For anyone building products on these models, this week moved real numbers. Multi agent coding loops that burned budgets on flagship pricing now run forty percent cheaper, with the option to route everyday traffic through models that cost a tenth of what the flagship did a year ago. The frontier is still the frontier. It simply stopped charging like a luxury. The teams that retest their routing assumptions this week will feel the difference in next month's invoice, which is exactly the point the labs were making with those price tags.
Quick answers
What is this story about?
On September 22, the AI industry got two flagship launches in a single afternoon, and both were really about the same thing. Anthropic shipped Claude Opus 5.5, the first model in its new 5.5 family. OpenAI answered with GPT 6 Sol and GPT 6 Luna, budget siblings of the GPT 6 Astra line. The press releases talked about benchmarks. The pricing pages talked louder.
Why does this story matter?
Sonnet 5.5 and Haiku 5.5 arrive in the coming weeks, carrying the same efficiency gains down the stack. For anyone building products on these models, this week moved real numbers. Multi agent coding loops that burned budgets on flagship pricing now run forty percent cheaper, with the option to route everyday traffic through models that cost a tenth of what the flagship did a year ago. The frontier is still the frontier. It simply stopped charging like a luxury. The teams that retest their routing assumptions this week will feel the difference in next month's invoice, which is exactly the point the labs were making with those price tags.
Sources
- Reuters on the Opus 5.5 launch
- Unite.AI on Opus 5.5 pricing and safeguards
- AIPost on the same day price war
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.