Analysis
OpenAI and Anthropic both announced cheaper frontier-adjacent models within about 90 minutes of each other on September 22, cutting API prices by as much as half just ten days after both companies' CEOs publicly called for the industry to slow down. Anthropic went first, releasing Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens -- a 20% cut from Opus 5's $5/$25 list price, with Anthropic saying real-world savings reach 40% because the model completes tasks in fewer tokens and generates output more than 30% faster. OpenAI followed with GPT-6 Sol, priced at $2 per million input tokens and $10 per million output, down from GPT-5.6 Sol's $4/$20, and GPT-6 Luna at $0.10/$0.50, down from $0.20/$1.20 -- both released in ChatGPT Work, Codex and the API the same day, according to TechCrunch and Fortune.
Two Labs, Ninety Minutes Apart
Both releases sit one rung below each lab's actual flagship. GPT-6 Sol and Luna are cheaper, faster siblings to GPT-6 Astra, built with what OpenAI says are the same training methods that improved Astra's professional work, coding and factuality. Claude Opus 5.5 is Anthropic's first Opus-tier model to match the performance of its higher Fable-tier flagship, Fable 5.1, while costing less than the Opus it replaces. Neither company shipped a new top-end model -- this is a price and efficiency release, not a capability jump, which matters for how it should be read against the slowdown pledge.
“Claude Opus 5.5 is Anthropic's first Opus-tier model to match the performance of its higher Fable-tier flagship, Fable 5.1, while costing less than the Opus it replaces.”
A Pledge, Then A Price War
The timing is not incidental. Anthropic CEO Dario Amodei published an essay on September 12 arguing the industry should deliberately slow the pace of frontier capability gains, citing the OpenAI agent swarm that breached Hugging Face in July as evidence a more capable successor could seize control of internet-connected systems within six to twelve months. OpenAI's Sam Altman and xAI's Elon Musk both said within hours they agreed. Ten days later, both labs' first releases since that pledge are competitive price cuts on production models -- pacing capability, on this evidence, has not meant pacing releases or competition for enterprise workloads.
The price war is playing out on two fronts at once: cheaper new models like Opus 5.5 and Sol/Luna, and outright cuts to existing flagship pricing, as businesses shift spend toward whichever lab offers the best cost-per-task. xAI's Grok 4.7, which Pulse covered at launch, and Google's Gemini line are the other two frontier competitors whose own pricing now sits under pressure from both moves; neither had announced a matching cut as of this writing. An economist quoted by Fortune framed it bluntly: OpenAI and Anthropic are now driving down the price of AI and, with it, their own near-term ability to profit from and reinvest in the next generation of models.
The Numbers, And What They Miss
A 50% list-price cut on GPT-6 Sol and a 20-to-40% effective cut on Opus 5.5 are large moves by the standards of enterprise software pricing, but they follow a now-familiar cadence: both labs have cut prices with nearly every model generation since 2024 as inference costs fall and competition intensifies. What's different this time is the backdrop -- a public commitment, from both CEOs, to compete less aggressively on capability -- landing during a week when both companies are competing exactly as aggressively as before on cost and distribution.
For AI-application startups, cheaper Sol, Luna and Opus 5.5 tokens mean real gross-margin relief on inference-heavy products, and for VCs underwriting those companies it's one more data point that the cost curve continues to bend down faster than most 2025-era unit-economics models assumed. It also means startups building thin wrappers around a single model's pricing advantage are exposed to exactly this kind of cut eroding their differentiation overnight.
What the release undercuts is the credibility of the pacing pledge itself, not the technology. Neither company has published a concrete definition of what "pacing the frontier" excludes -- efficiency and pricing moves, evidently, don't count -- and OpenAI's parallel announcement that it will let third-party groups conduct safety assessments during training and deployment is a process commitment, not a capability constraint. Critics of the original slowdown pledge argued it was a coordination signal more than a binding limit; a price war ten days later is the first real test of that skepticism.
The number worth tracking isn't the percentage cut -- it's whether Google, xAI or a Chinese lab like Alibaba matches it within the next two weeks, which would confirm this is now a structural price floor across the industry rather than a two-lab standoff.