Claude Opus 5.5 vs GPT-6 Sol pricing — exactly half, 90 minutes later
On September 22, 2026, Anthropic and OpenAI released new models roughly 90 minutes apart and opened an API price war. Anthropic priced Claude Opus 5.5 at 4 dollars per million input tokens and 20 dollars per million output tokens, 20 percent below Opus 5, and said it matches its top model Fable 5.1 on most tasks. OpenAI then priced GPT-6 Sol at 2 dollars input and 10 dollars output, exactly half of Opus 5.5 on both counts. Its lightweight GPT-6 Luna costs 0.10 dollars input and 0.50 dollars output. OpenAI said these are permanent list prices, not introductory discounts
The three lines
- Opus 5.5 — 4 dollars input and 20 dollars output per million tokens, 20% below Opus 5; Anthropic claims about 40% lower total cost
- GPT-6 Sol — 2 and 10 dollars, exactly half, released about 90 minutes later; GPT-6 Luna at 0.10 and 0.50 dollars
- Context — both companies are preparing to list; whether the prices reflect efficiency or investor subsidy is not yet clear
Key questions
- How much does Claude Opus 5.5 cost
- **4 dollars per million input tokens and 20 dollars per million output tokens**, which Anthropic says is **20 percent** below Opus 5. | Model | Input (per 1M tokens) | Output (per 1M tokens) | Note | |---|---|---|---| | Claude Opus 5.5 | **4 dollars** | **20 dollars** | 20% below Opus 5 | | GPT-6 Sol | **2 dollars** | **10 dollars** | exactly half of Opus 5.5 | | GPT-6 Luna | **0.10 dollars** | **0.50 dollars** | high-volume, lightweight work | **There is a second, separate claim about total cost.** Anthropic says Opus 5.5 needs fewer tokens to finish complex workflows, so the **all-in cost of running it is about 40 percent below Opus 5.** One report puts cache reads at 0.20 dollars per million tokens, a cut of roughly 60 percent. **The 20 percent is a price list; the 40 percent is the company measuring usage.**
- How do GPT-6 Sol and Luna differ
- **Sol is the mid-tier developer model; Luna is the high-volume lightweight one.** Both branch off GPT-6 Astra, released on September 3 (see 「GPT-6 Astra system card」). | | GPT-6 Sol | GPT-6 Luna | |---|---|---| | Built for | code review, debugging, complex developer tasks | summarisation and other bulk work | | Input | 2 dollars | 0.10 dollars | | Output | 10 dollars | 0.50 dollars | | Versus predecessor | **50%** below GPT-5.6 Sol | up to **58%** lower | | Company claim | half the error rate on factuality evaluations | — | **Luna is the eye-catching number.** At 0.50 dollars per million output tokens, commentators noted it undercuts even low-cost Chinese open-weight models. OpenAI credits **technical breakthroughs in inference processing and data caching.**
- Why did the price war start now
- **Two pressures met — buyers watching costs and sellers heading toward public listings.** | Factor | Detail | |---|---| | Buyers | CFOs have seen **sticker shock** on AI bills (Fortune); cost now outranks raw capability | | Sellers | both companies are reportedly **preparing IPOs** at valuations near 1 trillion dollars | | Tactic | pricing at exactly half a rival on launch day says there is no cost reason to choose anyone else | | Same period | Microsoft is discounting Copilot 30 to 50 percent for large contracts from October | **Fortune flagged a contradiction.** Days earlier, both labs had spoken of pacing the frontier. They answered that these are cost-efficiency models, **not frontier models** representing a major capability jump (see 「What a frontier model is」).
Same day, 90 minutes apart, exactly half the price. That is the short version of what happened in the AI API market on September 22, 2026.
1. The price list
Anthropic released Claude Opus 5.5 first. About 90 minutes later OpenAI released GPT-6 Sol and GPT-6 Luna.
| Model | Company | Input (per 1M tokens) | Output (per 1M tokens) | Change |
|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 4 dollars | 20 dollars | 20% below Opus 5 |
| GPT-6 Sol | OpenAI | 2 dollars | 10 dollars | 50% below GPT-5.6 Sol |
| GPT-6 Luna | OpenAI | 0.10 dollars | 0.50 dollars | up to 58% lower |
Sol is exactly half of Opus 5.5 on both input and output. OpenAI confirmed these are permanent list prices. Only a month earlier, on August 21, GPT-5.6 Sol output had been cut from 30 to 20 dollars; it has now halved again.
2. What each company claimed beyond price
| Company | Claim | Nature |
|---|---|---|
| Anthropic | matches Fable 5.1 on most tasks | internal measurement |
| Anthropic | about 40% lower total running cost thanks to fewer tokens | internal measurement |
| Anthropic | beats GPT-5.6 Sol on software benchmarks at about one-third the cost | internal measurement |
| Anthropic | 85% fewer containment-bypass attempts in internal testing | internal measurement |
| OpenAI | Sol halves its predecessor's factuality error rate | internal measurement |
| OpenAI | cuts come from inference and caching advances | explanation |
Every row is self-reported. No independent benchmarks were available at publication. Readers should also separate unit price from total cost: identical per-token prices can produce very different bills if one model uses more tokens per task.
3. Why now, and why half
Demand side. Fortune reports that CFOs have hit sticker shock on AI spending, and buyers now rank cost optimisation ahead of raw capability. Microsoft is discounting Copilot by 30 to 50 percent for large contracts starting in October.
Supply side. Both labs are reportedly preparing public listings at valuations near 1 trillion dollars. Undercutting a rival by exactly half on launch day is a vivid story for investors.
| Reading | Support | Objection |
|---|---|---|
| efficiency gains | inference and caching advances, per OpenAI | underlying costs are not disclosed |
| investor-funded market share | neither company is yet profitable | OpenAI calls the prices permanent |
WION put it well: cheap frontier AI is a real benefit now, but assuming it lasts is a different bet.
4. One contradiction
In mid-September both companies had endorsed pacing the frontier. Fortune noted the tension; both replied that these releases are cost-efficient models, not frontier leaps. The technique most associated with moving top-model capability into cheaper models is explained separately (see 「What model distillation is」).
5. What remains unconfirmed
- Launch date — most outlets say September 22, WION says September 24.
- Performance and total-cost claims are all company measurements.
- Whether prices are below cost cannot be judged; neither company discloses it.
- A Google pricing response had not been confirmed at publication.
Sources
- Techstrong.ai — OpenAI, Anthropic launch dueling models in API price war
- Fortune — What slowdown? OpenAI, Anthropic release dueling models as AI price wars heat up
- Seoul Economic Daily — Anthropic, OpenAI roll out cut-price AI models
- WION — OpenAI priced its new model at exactly half of Anthropic's
- The Founder's Wire — OpenAI and Anthropic cut prices 90 minutes apart