On Tuesday, September 22, OpenAI released GPT-6 Sol and GPT-6 Luna, and Anthropic released Claude Opus 5.5. Both companies led with lower prices.

Within three days each had also changed something a paying developer would want to know. OpenAI fixed a bug that had degraded how the new models read images. Anthropic started charging again for some requests its safety classifiers refuse before producing any output.

This record sets the two companies' own documents side by side. Anthropic is one of its two subjects, and Hugin's records are drafted with the help of Anthropic's Claude models. Anthropic's documents are held here to the same standard as OpenAI's.

The prices, as each company states them

OpenAI's API changelog gives "Standard pricing per 1M tokens for prompts with up to 272K input tokens": GPT-6 Sol at $2 input, $0.20 cached input and $10 output, and GPT-6 Luna at $0.10, $0.01 and $0.50. For longer prompts the pricing page lists a long-context band at $4/$15 for Sol and $0.20/$0.75 for Luna.

Anthropic's launch post: "Input and output tokens are $4 and $20 per million, 20% less than Opus 5." Cache reads are $0.20 per million. Its pricing page says models from Claude 4.6 on "include the full 1M token context window at standard pricing."

For subscribers, Anthropic adds two terms without numbers:

In addition to the price drop, we’re increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. We’re also providing subscription users a rate limit reset, which you can now save and use whenever you choose.

The post does not say how large the increase is. Anthropic's help article on limit resets, read the same day, adds conditions the post leaves out. "An unused reset expires on the day and time listed on the offer." And: "If you downgrade or cancel before using your reset, it's no longer available." The article does not say which expiry applies to this reset.

What half price is measured against

OpenAI's launch post describes "reducing API prices for Sol and Luna by 50% compared with their GPT‑5.6 promotional pricing." Its baseline for GPT-5.6 Sol is $4/$20. The pricing page, read September 27, lists gpt-5.6-sol at $4.00/$20.00 and says "GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026." That is a floor, not an end date.

When this desk read the same page on August 7, it listed gpt-5.6-sol at $5.00/$30.00 on the Standard tier. Against that earlier price, GPT-6 Sol is 60% lower on input and about 67% lower on output. OpenAI's changelog dates the $4/$20 promotional price to August 21, calling it "20% lower input pricing and 33% lower output pricing." Those percentages match the August 7 price.

A bug fix three days after launch

OpenAI's changelog carries this entry, dated Friday, September 25:

Fixed a bug in image encoding that degraded image understanding in GPT-6 Sol and GPT-6 Luna. This update improves results on visual tasks in the API and Codex, including computer use.

If your use cases involve image inputs, we recommend rerunning your evaluations and retrying workflows affected by the issue.

The entry does not say when the bug began, how much it degraded results, which requests it reached, or whether any credit is offered.

The launch post's computer-use result, GPT-6 Sol at 60.5% on OSWorld 2.0 offline, came from evaluations OpenAI says "were performed in our research environment or via our API". The changelog does not say whether those runs were affected. The ChatGPT release notes, read the same day, say nothing about the fix.

A refusal before any output is a charge again

On Thursday, September 24, Anthropic's platform release notes announced it is "resuming billing for refusals that arrive before any output" when the classifier category is bio, frontier_llm or reasoning_extraction, "the categories where we measure low volumes of false positives." The note adds: "Mid-stream refusals were already billed." A June 2, 2026 entry on the same page had said that on the Claude API "you are no longer billed for a request" refused "without Claude having generated any output." The refusals documentation sets out the terms:

These refusals are billed like any other request, at the rates of the model that ran it. A refusal before any output in any other category, or with a null category, is not billed. Either way, content is empty and token counts appear in usage. The request still counts against your rate limits.

The page says these charges exist "To disrupt attempts to circumvent Anthropic's safeguards at scale". Its category table says legitimate work can trigger two of the three billed categories. For bio: "Beneficial life sciences work can also trigger this category." For frontier_llm: "Benign machine learning work can also trigger this category." The page publishes no false-positive figures, and it keeps the list open: "The billed categories may change as Anthropic keeps measuring and refining its safeguards' false positive rates."

The help article for Claude app users states the rule in its own words, naming the biology, distillation and frontier LLM development classifiers, and warns that the classifiers read "memory, content from connectors, web search results, and files, so a fallback can be triggered by content you didn't type." It does not say how a billed refusal counts against a subscription's usage limits.

The model that answers can cost more than the one you chose

Anthropic's launch post says users "will be able to identify and fix bugs in their code as part of the routine software development lifecycle, but most cybersecurity tasks will be re-routed to Opus 4.8." Its help article says biology and frontier-LLM-development flags on Opus 5.5 fall back to Opus 5, and that for distillation "the request is blocked outright."

Anthropic's pricing page lists both fallback models, Opus 5 and Opus 4.8, at $5/$25 per million tokens and $0.50 for cache reads. Opus 5.5 is $4/$20 and $0.20. Where those are the fallbacks, a flagged request is answered at a higher per-token price than the model the developer selected. For the billed categories there is a second charge: when fallback is used, "the refusal that triggered it is billed, in addition to the fallback request". A fallback credit covers the fallback's prompt-cache miss.

On the API, fallback is off until configured; the help article says "API customers must opt into and configure the fallbacks." Without it, a flagged request comes back as an ordinary 200 response whose stop reason is a refusal. With it, the refusals documentation says default routing "is not published per model on the Models API"; the response names the model that answered.

The routing reaches Anthropic's own benchmark table. Its footnote says Opus 5.5 "was evaluated with its production safeguards enabled. When they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5." OpenAI's post says its AutomationBench figure for Claude Fable 5.1 "understates its actual cost, as it omits the cost of the Opus 5 fallbacks, which occurred on ~40% of tasks." That is one company's claim about a competitor, and this desk has not tested it.

What to check

Rerun image work. If you ran GPT-6 Sol or Luna on images, screenshots or computer use before September 25, OpenAI recommends rerunning.
Log Anthropic refusal categories. Since September 24, a pre-output refusal in bio, frontier_llm or reasoning_extraction is billed; one in cyber or general_harms is not.
Read the per-attempt record. With fallback on, Anthropic calls usage.iterations "the per-attempt record of what you're billed."
Check a saved reset's expiry in the usage settings. The help article says the reset button is not in Claude Mobile or in Claude Code in a terminal or IDE.
Migrating to Opus 5.5: per the release notes, setting thinking to disabled or enabled, or tool_choice to any or tool, returns a 400 error. Anthropic's deprecations page lists its retirement as not sooner than September 22, 2027.

What the record does not say

When OpenAI's image bug began, how large the degradation was, or whether any credit is offered.
Whether OpenAI's launch benchmarks ran with the bug.
Why Anthropic stopped billing pre-output refusals on June 2, or how many refusals the September 24 change now bills.
Anthropic's false-positive rates for any category.
The size of the five-hour limit increase, or this reset's expiry.
Whether OpenAI bills requests stopped by its own safety checks. Its two safety-check pages, read the same day, do not say either way.
Every benchmark and cost comparison in both launch posts is the vendor's own.