September 2026 AI Pricing Roundup: GPT-6, Opus 5.5 and Cheaper Flash Models
Nine new models in four weeks. Most cost the same as, or less than, the models they replaced; GPT-6 Astra and DeepSeek V4.1 Flash are the exceptions.
The month in one table
| Date | Model | Input / output per 1M | What changed |
|---|---|---|---|
| 1 Sep | Claude Fable 5.1 | $10.00 / $50.00 | Replaced Fable 5 at the same price; cached input cut from $1 to $0.25 |
| 2 Sep | Gemini 3.8 Flash | $0.75 / $3.75 | Promotional price until 31 December, then $1.50 / $7.50 |
| 2 Sep | Muse Spark 1.3 | $1.25 / $4.25 | Meta's new frontier model |
| 3 Sep | GPT-6 Astra | $10.00 / $50.00 | OpenAI's most expensive standard model |
| 10 Sep | DeepSeek V4.1 Flash | $0.30 / $1.20 | Replaced V4 Flash ($0.14 / $0.28); half price off-peak |
| 21 Sep | Grok 4.7 | $2.00 / $6.00 | Same price as Grok 4.6 for prompts under 200K tokens |
| 22 Sep | Claude Opus 5.5 | $4.00 / $20.00 | 20% below Opus 5 |
| 22 Sep | GPT-6 Sol and Luna | $2.00 / $10.00, $0.10 / $0.50 | Replaced GPT-5.6 Terra and Luna at lower prices |
Prices kept falling
Our token price index tracks the median blended price of current models in each tier. In September the frontier median fell from $4.50 to $3.75 per 1M tokens, as Meta's Muse Spark 1.3 joined the tier at $1.25 / $4.25 and Opus 5.5 replaced the pricier Opus 5. GPT-6 Astra, at $10 / $50, pulled the other way. On 27 September it stood 9% below its level on 13 June. The workhorse and small-model medians held steady this month, after falling by about a third over the summer; most of the workhorse drop came in July, when Google launched Gemini 3.6 Flash at half the price of 3.5 Flash.
OpenAI: GPT-6 arrives in three sizes
GPT-6 Astra, at $10 / $50, sits at the top of OpenAI's range, while GPT-6 Sol ($2 / $10) and GPT-6 Luna ($0.10 / $0.50) cost less than the GPT-5.6 Terra and Luna models they replace. Luna is the cheapest current-generation OpenAI model, at half GPT-5.6 Luna's input price. All three share a 1.05M-token context window, with double input pricing above 272K tokens. Full GPT-6 pricing.
Anthropic: a cheaper Opus, a cheaper cache for Fable
Claude Fable 5.1 kept Fable 5's $10 / $50 but cut the cache-read price by 75%, from $1 to $0.25 per million tokens, which lowers the cost of long agent sessions considerably. Three weeks later, Claude Opus 5.5 arrived at $4 / $20 with cached input at $0.20. On the Arena text leaderboard it posted a provisional 1509, ahead of Fable 5.1's 1501. Full Opus 5.5 pricing.
Google, xAI and DeepSeek
Gemini 3.8 Flash is the third Flash release since July at the same promotional $0.75 / $3.75. Google has said the price doubles on 1 January, so budgets for next year should use $1.50 / $7.50. Grok 4.7 kept xAI's $2 / $6 price. DeepSeek V4.1 Flash replaced V4 Flash and costs more than its predecessor, $0.30 / $1.20 at peak against $0.14 / $0.28, though its off-peak rate of $0.15 / $0.60 is close to the old price and its cache hits cost just $0.006 per million tokens.
What to do now
- On Opus 5 or older Opus models: test Opus 5.5. It is cheaper on every rate.
- On GPT-5.6 Terra or Luna: GPT-6 Sol and Luna cost less. Check the long-context surcharge if your prompts exceed 272K tokens.
- On Gemini Flash: plan for the price to double on 1 January 2027.
- On DeepSeek V4 Flash: your requests now go to V4.1 Flash at the higher price. Move batch work to off-peak hours.
Coming up
- 23 October: OpenAI shuts down o1 and the gpt-4o-2024-05-13 snapshot.
- 21 November: GPT-5.6 Sol's promotional $4 / $20 is guaranteed until at least this date.
- 11 December: OpenAI shuts down the o3-2025-04-16 snapshot.
- 31 December: Gemini 3.8 Flash's promotional price ends.
Every change is logged with its source in our price history, and the RSS feed brings new ones to your reader. To see what this month's changes mean for your own bill, use the token cost calculator.