Sol costs half as much as Claude Opus 5.5 — but Anthropic holds its price all the way to one million tokens
On Tuesday, September 22, 2026, OpenAI launched two new models in the GPT-6 family — Sol and Luna — which the company prices at roughly half of their GPT-5.6 predecessors. The same day, Anthropic released Claude Opus 5.5, which the company says costs 40 percent less for typical work tasks. Both price savings are the companies' own claims, but the price tables now published show a concrete competition in which the price per token has become the most measurable battleground between the two companies.
What the models are
GPT-6 Sol and Luna were launched on Tuesday, September 22, and are marketed as lighter and cheaper variants of GPT-6 Astra, the large model OpenAI released earlier in the month, writes Gizmodo (Gizmodo). Gizmodo describes them as "mini-Astras," and they are priced against the GPT-5.6 generation.
OpenAI's announcement states that the models were trained with "similar methods as GPT-6 Astra" — which could include recurrent depth, a technique Gizmodo describes as contested because it can boost performance while making it harder for researchers to follow the model's decision processes. OpenAI has not confirmed whether recurrent depth was actually used, and the company did not respond to Gizmodo's inquiry on the matter.
Technical specifications compiled by TechRepublic give Sol a context window of 1.05 million tokens — Opus 5.5 offers standard pricing through a window of one million tokens — and a wider range of reasoning controls: Sol can run with reasoning switched off entirely ("none"), in addition to low, medium, high, xhigh and max, while Opus 5.5 only goes from low to max.
The numbers behind the comparison
Standard API price per million tokens, according to TechRepublic's compilation (TechRepublic):
| Category | GPT-6 Sol | Claude Opus 5.5 |
|---|---|---|
| Input (standard) | $2 | $4 |
| Output (standard) | $10 | $20 |
| Cache reads | $0.20 | $0.20 |
| Long context | Higher rates above 272,000 input tokens | Standard rate all the way up to 1 million tokens |
At standard rates, Sol is half the price of Opus 5.5 on both input and output, while cache reads are equal. But once the context grows long, the picture flips: Sol's rates rise above 272,000 input tokens, while Opus 5.5 keeps its standard rates through the entire million-token window — a point that can matter a great deal for agents working with large codebases or long documents.
Luna is the lightest model. According to a review on note.com citing OpenAI's announcement (note.com), the API price is $0.10 per million input tokens and $0.50 for output — stated as a 50 percent drop from GPT-5.6 Luna's promotional price. The same source says Luna is available to free and Go users via the desktop app, while Sol ($2/$10) requires Plus or higher on ChatGPT Work and Codex plans and is not available in standard Chat. Note that this source is a machine-translated analysis; the pricing and availability details should be cross-checked against OpenAI's primary documentation.
Where each model wins — and loses
TechRepublic's early comparison concludes that GPT-6 Sol starts at half of Opus 5.5's standard price, but that Claude delivers stronger results on several current coding benchmarks. The difference changes, however, when the prompts become very large — where Sol's larger context window and Opus 5.5's standard pricing across the whole window come into play. The models also differ in reasoning controls, deployment options and safeguards, but the sources are truncated on precisely these points, so the details are not yet fully documented.
As for deployment: Sol runs in OpenAI's Codex, Opus 5.5 in Claude Code.
Market reaction: the agent economy
Box CEO Aaron Levie commented on the launches on X on Tuesday, as reported by Yahoo Finance/Benzinga (Yahoo Finance). As reported there, he pointed out that the frontier models had just become significantly cheaper — with the Opus 5.5 price cuts and then GPT-6 Sol and Luna, which according to him lowered token prices by 50 percent — and that falling AI costs dramatically increase the number of economically viable agent use cases. This is Levie's own assessment; we report it but cannot verify it as market analysis.
Yahoo Finance/Benzinga sums up the same two-sided picture: OpenAI launched Sol and Luna with API prices stated as up to 50 percent lower than GPT-5.6's promotional pricing, while Anthropic reduced the cost of its newest model.
What is verified, and what is a company claim
Documented through published specifications and corroborating sources: the launch date (Tuesday, September 22, 2026), the API price tables above, the context windows and the reasoning levels. Note that even this rests on secondary reporting quoting the companies' announcements — none of the sources in this evidence base is a primary press release, so the exact official wording cannot be verified first-hand.
Company claims not independently confirmed: Anthropic's assertion that Opus 5.5 costs 40 percent less for typical work tasks, OpenAI's "up to 50 percent lower" framing for Sol and Luna compared with GPT-5.6, and Levie's characterization of 50 percent lower token prices.
Open questions
The biggest is recurrent depth: OpenAI said the models were trained with "similar methods as GPT-6 Astra," but has not confirmed whether that includes recurrent depth — the technique Gizmodo describes as controversial precisely because it can make the model's decision processes harder to monitor. Second, the sources lack complete details on safeguards and deployment. And third, the coding-performance comparison so far rests on TechRepublic's qualitative assessment; we lack benchmark numbers showing how Sol and Opus 5.5 actually stack up against each other over time and across different workloads.
What is nonetheless clear on the same day: both companies chose to compete on price, not just performance — and that makes cost per token one of the frontier race's most measurable axes.