AI & Machine Learning
Frontier AI got 20% cheaper in September. Budget for the meter anyway.
6 min read
Claude Opus 5.5 cut frontier list prices by a fifth, but introductory prices expire, the best tools are gated and agents are moving onto meters. Why AI bills can rise while prices fall, and how to budget for 2027.
In brief
September's model releases pushed top-tier list prices down. On 22 September Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Anthropic says it matches the company's larger Fable 5.1 model "on most work". OpenAI added GPT-6 Sol and Luna the same day. An aggregator reports they cost half as much, but ByteNib could not confirm that on OpenAI's own site.
Four other developments that month point the other way for anyone setting a 2027 AI budget. Introductory prices expire. The strongest capabilities now sit behind vetting programmes. AI features at work and at home are moving onto usage meters. And models are being retired fast enough that switching is now a recurring cost.
The price moves
| Date | Model | What changed | How well sourced |
|---|---|---|---|
| 1 Sep | Anthropic Fable 5.1 and Mythos 5.1 | The same model under two levels of safeguards. Mythos 5.1 is limited to vetted US organisations. Fable 5.1 lists at $10 and $50; a 75% cut in the price of cached input makes it about 25% cheaper to run than Fable 5 on typical workloads. | Company pages |
| 2 Sep | Google Gemini 3.8 Flash and Flash Cyber | Flash costs $0.75 and $3.75 until 31 December, then $1.50 and $7.50 from 1 January 2027. Flash Cyber is open only to "trusted defenders". | Company page |
| 3 Sep | OpenAI GPT-6 Astra | $10 and $50, doubled for prompts over 272,000 tokens. It is OpenAI's first model rated "Critical" for cybersecurity, and it is off by default in enterprise workspaces. | OpenAI developer docs |
| 10 Sep | DeepSeek V4.1 Flash | Retired model names now point to the new model, with prices "reduced accordingly". | DeepSeek change log |
| 22 Sep | Anthropic Claude Opus 5.5 | $4 and $20, from $5 and $25. Cached input costs $0.20, from $0.50. A faster mode costs $8 and $40. Anthropic says typical workloads cost about 40% less to run than on Opus 5. | Company page; benchmarks company-reported |
| 22 Sep | OpenAI GPT-6 Sol and Luna | Confirmed in OpenAI's own Codex release notes. The "50% lower API prices" claim comes only from an aggregator. | Partly verified |
Prices are per million input and output tokens at list price. Enterprise agreements, committed-use discounts and cloud-marketplace pricing can differ.
Why the bill can rise while prices fall
Introductory prices end. Gemini 3.8 Flash doubles in price on 1 January 2027. A 2027 budget built on today's rate will be short by half on every workload that runs on it, and the change arrives on the first day of most organisations' budget year.
Agents use more than chat. A cheaper token matters less when the workload multiplies. Agents that plan, call tools, retry and check their own work consume far more tokens per task than a chat reply. Vendors are pricing for that. On 25 September Microsoft moved its new Copilot's more advanced agent features to usage-based billing without publishing a rate card. Consumer products followed the same pattern. Apple says it will add usage limits to its new Siri and that "increased access to such features will be available for a fee". From November, Xbox Cloud Gaming will cap monthly play hours on its subscriptions and sell extra time.
The best tools come with a vetting step. All three large US labs now ship their strongest cyber or biology capabilities in a gated tier. Anthropic's Mythos 5.1 goes through its Cyber Verification and Life Sciences Verification programmes. Google's Flash Cyber goes through its Fairwind programme, which is open to government bodies, critical-infrastructure operators and software maintainers. OpenAI's Daybreak programme is meant to relax Astra's limits for vetted defenders. What a security team can use now depends on whether a vendor approves it, not only on price.
Swapping models is a recurring cost. Opus 5.5 arrived two months after Opus 5, and Gemini 3.8 Flash came three weeks after Google's previous release, according to 9to5Google. Retirements forced moves too. OpenAI shut the Sora API on 24 September. On 18 September Google restricted Gemini 2.5 to existing users, and a Google preview model shuts on 5 October. Every change needs testing before production traffic moves.
What leaders should do next
- Budget on cost per task, not cost per token. Measure tokens per completed task on your own workloads. A cheaper model that needs more retries or longer answers can cost more in practice.
- Put expiry dates in the budget. List every introductory or promotional price you rely on, with the date it ends. Gemini 3.8 Flash's 1 January rise is the first large one.
- Give model lifecycles an owner. Keep a calendar of retirement dates, fix model versions in configuration where the vendor allows it, and treat silent re-routing as a change that needs testing.
- Decide on gated tiers deliberately. If your security or research teams want the stronger models, start the vetting applications now, and plan for work continuing if access is withdrawn.
- Cap the meters before they start. Set spending limits and approval rules for usage-billed agents before anyone switches them on. ByteNib's guide to Copilot's move to usage billing sets out the dates.
Sources and scope
Model details and prices come from Anthropic on Opus 5.5, Anthropic on Fable and Mythos 5.1, Google on Gemini 3.8 Flash, OpenAI's developer documentation for GPT-6 Astra, DeepSeek's change log and OpenAI's Codex release notes. TechCrunch reported on Opus 5.5 and CSO Online on Astra's enterprise rollout. Retirements are from OpenAI's help centre and the Gemini API changelog.
The metered-pricing examples are from Microsoft, MacRumors and Xbox Wire. OpenAI's pages could not be read directly for GPT-6 Sol and Luna pricing, so that claim is reported as unverified. Benchmark results are the companies' own and were not independently tested.
The budgeting analysis and recommendations are ByteNib editorial analysis. This article was researched and drafted with AI assistance, including Anthropic's Claude, and edited under ByteNib's editorial standards. Anthropic is one of the vendors assessed here.
Continue exploring: AI & Machine Learning analysis, the related implementation tutorial, and the structured learning path.