OpenAI cut the price of its GPT-6 Sol model by half in September. Anthropic did the same thing a week later with Claude Opus 5.5, pricing input tokens at $4 per million against the $5 Opus 5 charged. Google closed the month with Gemini 4 Argon at $2 per million input tokens.
The numbers tell the story. OpenAI's Luna model, built for high-volume work like summarizing documents, now costs $0.10 per million input tokens and $0.50 per million output tokens. Anthropic says Opus 5.5 runs about 40% cheaper than Opus 5 on typical workloads and generates output more than 30% faster. Cache reads on Opus 5.5 dropped 60%, to $0.20 per million tokens. Google is offering a 95% discount on cached input for Gemini 4 Argon during the introductory period.
The price cuts matter now because AI spending is under pressure. Companies are watching their API bills climb while their returns stay fuzzy. The Economist reported this week that businesses are routing work through brokers like OpenRouter, which let them swap between labs when one is offering a bargain. CNBC reported that OpenAI and Anthropic are both fighting cheaper open-weight models from Alibaba, Moonshot AI and DeepSeek. Anthropic's own listing page notes that a popular DeepSeek model sells at roughly one-fortieth the price per token of a frontier Anthropic model.
The pricing war nobody expected
The three labs have converged on almost identical rates for their mid-tier models. Sol, Opus 5.5 and Gemini 4 Argon all start at $2 input and $10 output per million tokens during their introductory windows. A year ago that tier cost roughly double.
"One of the things we're continuing to innovate on is how to make that thinking, how to make the answering more efficient, so it uses less tokens depending on your effort setting," Dianne Penn, head of product management at Anthropic, told CNBC. That sentence describes the whole game now. The labs aren't only racing on raw capability. They're racing on how few tokens a task eats.
Anthropic published the math. Opus 5.5 at its default effort level beats GPT-6 Astra's top score on the FrontierCode benchmark for about a fifth of the cost per task. On Terminal-Bench 4.0 it matches Astra for roughly 40% of the cost. On CursorBench it beats GPT-5.6 Sol's top score by 11 points at about a third of the price.
OpenAI's answer came a week after Anthropic's release. GPT-6.1 Sol landed on 29 September at the same base rates as GPT-6 Sol, with cached input at just $0.10 per million tokens. OpenAI says it matches Astra on the DeepSWE v1.1 software engineering benchmark and scores 2.2 percentage points above Opus 5.5 on AutomationBench, at a third of the cost.
Those are the labs' own numbers, run on the labs' own benchmarks. Take them with the expected salt. Nobody outside either company can reproduce a 40% cost advantage claim without building the same harness.
Where the real cost of AI work is hiding
The biggest price move isn't in the headline rates. It's in cache reads.
A cache read is when the model reuses context it has already processed, the same long system prompt or the same 200,000-line codebase it read on the previous turn. For agentic coding work, cache reads make up most of the bill. Anthropic cut cache reads from $0.50 to $0.20 per million tokens, a 60% drop. OpenAI cut cached input on GPT-6.1 Sol to $0.10 per million, a fifth of Astra's standard rate. Google is discounting cached input by 95% on Gemini 4 Argon.
That's where the savings land. A company running a coding agent across a big repository sees most of its tokens flow through cache reads. The headline input price barely moves the bill. The cache price decides whether the agent is affordable.
Anthropic's Opus 5.5 page makes this concrete. An early tester audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5 times as many tokens. Another tester finished a 680,000-line code migration in less than a day. "It writes the way I do," that tester said.
GitHub's chief product officer Mario Rodriguez put it in terms his users would recognize: "In our testing across GitHub Copilot CLI and VS Code, Claude Opus 5.5 used among the fewest tokens and steps we measured. In VS Code, it solved more terminal tasks than Opus 5 in less than half the steps."
Whether that holds across a normal week of ordinary work, not a demo migration, is unclear. Lab benchmarks measure specific, repeatable tasks. Real engineering is messier. It's not obvious the efficiency gains scale cleanly.
The cybersecurity tier that isn't for everyone
There's a quieter shift in these releases, and it may matter more than the price cuts in two years.
OpenAI classifies GPT-6 Astra as the first model in the GPT-6 family to reach what the company calls the "Critical" threshold on its internal cybersecurity scale. The model scored 100% on exploit development benchmarks in internal tests and found two previously unknown zero-day vulnerabilities. GPT-6.1 Sol carries the same classification.
Anthropic is doing something similar. Opus 5.5 routes most cybersecurity tasks to Opus 4.8 by default. Verified cyberdefenders can apply to the Cyber Verification Program, which has three tiers of trusted access. Biology work is handled the same way, through a new Life Sciences Verification Program for vetted labs and pharmaceutical companies.
Google closed the loop on 30 September. Gemini 4 Argon scores 68% on CWE-bench v1, a benchmark for fixing security vulnerabilities, which puts it joint first. Initial access is limited to verified cybersecurity defenders through Google's Fairwind Program. Koray Kavukcuoglu, SVP at Google DeepMind, said a phased release approach is needed for a model with that capability profile.
That's a new pattern for this industry. The best models on the market are gated. A developer who wants Opus 5.5 for ordinary software work gets it today. A developer who wants to use it for advanced penetration testing has to apply and wait.
The argument is straightforward on paper. A model that writes working exploits faster than any human team is a real risk if it's freely available. The labs decided to treat that risk the way they treat biological work: verify the user, log the access, restrict the capability.
It's less clear how long the gating holds. Google's introductory price on Gemini 4 Argon rises after the promotion window, from $2 to $4 per million input tokens. The same logic could apply to access. As the model gets older and the security profile becomes routine, the walls come down.
What's driving the price drop
The labs cite efficiency, and that's part of it. Dianne Penn at Anthropic described the shift as using fewer tokens per task. The token efficiency gains are real. They show up in the cost-per-task numbers.
But there's a second force. Cheaper open-weight models from China are taking share. Alibaba, Moonshot AI and DeepSeek are all shipping models that are good enough for a lot of production work and cost a fraction as much. Forbes reported this week that startups are emerging to help businesses use AI more cheaply, in part by shifting to open Chinese models.
That pressure is what turns an efficiency gain into a permanent price cut. A lab that could charge $5 per million input tokens a year ago is matching $2 today because the alternative is losing the customer.
Anthropic's pricing history shows the pattern. Fable 5.1 kept the same $10 input rate as its predecessor but cut cache reads by 75%. Opus 5.5 cut base input by 20% and cache reads by another 60%. Sonnet 5.5 held its rates but raised speed and cut cost per task by up to 30%. Every release in the family is cheaper per unit of work than the one before.
OpenAI did the same. Reuters reported in August that the company cut GPT-5.6 Sol pricing by more than 20% for three months. The September releases made that structure permanent.
What buyers should watch
The gating is the part to watch. Google's Koray Kavukcuoglu described a phased approach for Argon. That phrase covers a lot of possibilities, from a short verification queue to an indefinite closed program. If the walls stay up, companies running security work will end up routing it to older, unrestricted models, which is exactly what Anthropic is doing with Opus 4.8.
That has a cost. The unrestricted model is slower and weaker. A team that needs the latest capability for defensive security work is stuck with a worse model or a waiting list. Anthropic's own benchmark footnote admits the point: when cybersecurity safeguards intervened during evaluation, tasks fell back to Opus 4.8, and "this likely reduces Claude Opus 5.5's performance on these benchmarks."
The tradeoff is defensible. It's also a decision that the labs are making on behalf of their customers, without much of a public process. If the gates stay up, the pricing war only applies to the part of the market that isn't doing security work.
Where this goes next
OpenAI released three models in September and Anthropic released three. Google released one, gated. Sonnet 5.5 and Haiku 5.5 are still coming from Anthropic, with more of the same efficiency and safety changes.
The trend is structural. Every tier is getting cheaper per unit of real work, not just per token. Cache discounts are doing most of the heavy lifting for coding agents, which is where most of the money is going.
For buyers, the math is simple. If you're running agents on big codebases, the cache price is the number that matters. If you're doing high-volume text work, the Luna class of models is now near-free. If you need security work, you're waiting for verification.
None of that is settled. The gating could loosen as the models age. The efficiency gains could prove narrower in real use than on lab benchmarks. The Chinese models could take more share than either American lab expects. What won't reverse is the direction: each generation is cheaper per unit of real work than the last.
More on AI pricing and model releases in our AI coverage.