AI API price cuts arrived from OpenAI and Anthropic within hours of each other on 22 September 2026. OpenAI halved rates for its new GPT-6 Sol and Luna models, while Anthropic cut Claude Opus 5.5 by 20% against Opus 5. All the figures below are US dollar list prices.
List prices are what the vendors publish for pay-as-you-go API use. Large enterprise agreements are negotiated separately and may differ, and none of these rates applies to consumer chat subscriptions. The cuts take effect with the models themselves, which are available now.
What Each Model Now Costs
Both companies price by the million tokens, a token being roughly three-quarters of an English word. Input tokens are what you send the model; output tokens are what it returns, and output is always dearer.
| Model | New input / output per million tokens | Previous | Change |
|---|---|---|---|
| GPT-6 Sol (OpenAI) | $2 / $10 | $4 / $20 (GPT-5.6 Sol) | 50% lower on both |
| GPT-6 Luna (OpenAI) | $0.10 / $0.50 | $0.20 / $1.20 (GPT-5.6 Luna) | 50% input, 58% output |
| Claude Opus 5.5 (Anthropic) | $4 / $20 | $5 / $25 (Claude Opus 5) | 20% lower on both |
| GPT-6 Astra (OpenAI flagship, 3 September 2026) | $10 / $50 | Unchanged | No cut |
As of 22 September 2026, the practical effect is that Claude Opus 5.5 and GPT-6 Sol sit at very different points: Opus 5.5 matches the old GPT-5.6 Sol price exactly, at $4 and $20, while the new Sol undercuts it by half. The two flagship tiers, GPT-6 Astra and Anthropic’s Fable 5.1, both remain at $10 and $50 and were not reduced.
The Biggest Cut Is on Cached Input
Prompt caching lets a customer store a block of context, such as a long document or a system prompt, and reuse it across calls at a lower rate than sending it fresh each time. For workloads that reread the same material repeatedly, which is most agent work, this line is often the largest part of the bill.
- Claude Opus 5.5 cache reads: $0.20 per million tokens, down from $0.50 on Opus 5, a 60% reduction.
- Claude Opus 5.5 cache writes: $5 per million for the five-minute tier, down from $6.25.
- GPT-6 Sol cached input: $0.20 per million, with cache writes at $2.50.
- GPT-6 Luna cached input: $0.01 per million.
Anthropic’s cache-read cut is proportionally three times larger than its headline cut. OpenAI has said GPT-6 improves cache hit rates by default so that agents reuse more context automatically.
OpenAI Says the New Rates Are Permanent
The obvious question about a 50% cut is whether it lasts. OpenAI has answered it directly: a company spokesperson told VentureBeat that “these GPT-6 Sol and Luna rates are permanent prices, not promotional or introductory pricing”.
That statement matters because the previous GPT-5.6 rates had been described in some coverage as promotional. A permanent list price can still be changed by the vendor at any time; the commitment is that these are not time-limited introductory rates due to expire on a set date.
Anthropic has made no equivalent statement about the duration of the Opus 5.5 pricing. It has published a model retirement commitment instead, undertaking not to withdraw Opus 5.5 before 22 September 2027.
Anthropic’s Cut Comes With Four Breaking Changes
A price cut on a new model version is only a saving if the existing code runs on it. According to the technical publication Digital Applied, moving from Opus 5 to Opus 5.5 involves four API changes that will return HTTP 400 errors on unmodified code.
- Thinking cannot be disabled: callers must use adaptive mode or omit the setting.
- Forced tool use removed: the tool_choice: “any” parameter must be replaced with structured outputs.
- Thinking blocks cannot be replayed: not after a system prompt or tool definition changes.
- Deprecated computer-use tool: the older tool identifier must be migrated to its 2026 replacement.
These migration requirements appear in that publication’s analysis rather than in Anthropic’s own launch announcement, and the list should be checked against Anthropic’s API documentation before anyone plans a migration. The point for a buyer is that the 20% saving carries engineering work, whereas OpenAI’s cut applies to newly released models rather than requiring an in-place upgrade of existing ones.
How Anthropic Gets to 40%
Anthropic’s own headline figure is not 20%. In its announcement of Claude Opus 5.5, the company said that “at default settings it will cost 40% less than Opus 5 on typical workloads”.
Both numbers are real and they measure different things. The 20% is the change in the published per-token rate. The 40% is a modelled cost for a whole workload, which also captures the cheaper cache reads and the fact that the new model reaches an answer using fewer tokens. Digital Applied noted that the comparison is drawn at different reasoning effort levels, medium for Opus 5.5 against high for Opus 5, so the workloads being compared are not identical.
Anthropic said it is “passing these efficiency savings on to our customers in the form of price cuts and rate limit increases”, and separately that it is increasing five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. Claude Opus 5.5 is available through the Claude Platform and on Amazon Web Services, Google Cloud and Microsoft Azure.
What This Does Not Change
The cuts are real but narrow, and several things a finance team might hope for are not in them.
- Flagship pricing: the top tiers at both companies stayed at $10 and $50 per million tokens.
- Seat prices: these are API rates. Per-seat business and enterprise chat subscriptions were not reduced.
- Total spend: a lower unit price reduces a bill only if usage holds steady. Industry reporting through 2026 has repeatedly described enterprises whose token consumption outran their forecasts.
- Negotiated contracts: customers on committed-spend agreements do not automatically inherit list-price changes.
- Comparability: two models at the same posted rate can cost very different amounts per completed task, depending on how many tokens each consumes.
The same week saw price movement elsewhere in the market. readus247 covered StepFun’s Step 5 Preview at $1 a million tokens on 21 September and DeepSeek’s V4.1 Flash price cut earlier in the month.
Frequently Asked Questions
What Are the New AI API Price Cuts?
On 22 September 2026 OpenAI released GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and GPT-6 Luna at $0.10 and $0.50, each about half the corresponding GPT-5.6 rate. Anthropic released Claude Opus 5.5 at $4 and $20, down 20% from Claude Opus 5.
Is GPT-6 Sol Cheaper Than Claude Opus 5.5?
On posted list prices, yes. GPT-6 Sol is $2 input and $10 output per million tokens against $4 and $20 for Claude Opus 5.5. Actual cost per task also depends on how many tokens each model uses, which the list price does not capture.
Are OpenAI’s Lower Prices Temporary?
An OpenAI spokesperson said the GPT-6 Sol and Luna rates are permanent prices rather than promotional or introductory pricing. A published list price can still be revised later.
Did the Flagship Models Get Cheaper?
No. GPT-6 Astra and Claude Fable 5.1 both remained at $10 per million input tokens and $50 per million output tokens as of 22 September 2026.
Why Does Anthropic Say 40% When the Rate Fell 20%?
The 20% is the per-token list price change. The 40% is Anthropic’s modelled saving on a typical workload at default settings, which also reflects a 60% cut to cache reads and fewer tokens used per task.
Do These Rates Apply to Business Chat Subscriptions?
No. These are application programming interface rates charged by usage. Per-seat business and enterprise chat plans are priced separately and were not cut.
