Six OpenAI models retire on December 11, and the suggested replacement costs 6.8×

The December 11 cliff

OpenAI shuts down gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro on December 11, 2026. The deprecation page maps each one to a GPT-5.6 model. Follow that page literally and your bill goes up several times over on the day.

Same workload on each row: 500M input tokens, 100M output, no caching.

Two reasons the named replacement is the wrong destination. The list was written before GPT-6.1 Sol and GPT-6 Luna existed, and both are cheaper than the models they stand in for. And 5.6 Sol's $4/$20 is a promotion OpenAI only promises “at least through November 21” — three weeks after it becomes the default replacement, nobody has published what it costs then.

The gpt-5-mini row is the one that stopped us. $325 becomes $2,200 — 6.8× — if you trust the link. GPT-6 Luna does the same job for $100 and scores 38 on Artificial Analysis at max effort against 5.6 Luna's 37. It is not a like-for-like swap, and a mini-tier workload is exactly the kind you would test on the cheap model before paying for Terra.

A 90% cut, with a condition

Anthropic released Claude Haiku 5.5 on October 7 at $0.10/$0.50 per million — a tenth of Haiku 4.5's $1/$5, and the same per-token price as GPT-6 Luna. The catch is further down the announcement: a request whose prompt exceeds 100,000 tokens is billed at five times the rate. If your context runs long, the headline number does not apply to you.

The change that did not get headlines is the cached input rate. Anthropic halved it on Sonnet 5.5, from $0.20 to $0.10. With GPT-6.1 Sol's cache also at $0.10, both vendors' main mid-priced models now sit at $2/$10 with a $0.10 cache.

For agent workloads the cache price matters more than the headline. An agent that resends the same context on every step pays for those input tokens over and over; Anthropic puts the saving from the halved cache at about 20% for that shape of workload — roughly $369 a month instead of $430, on 750M input tokens with 85% cached.

One caveat on Sonnet itself, because rate card comparisons flatter newer models. Sonnet 4.5 retires November 30 and its replacement Sonnet 5.5 is cheaper on paper ($3,000 to $2,000 on the same workload), but Anthropic's newer tokenizer counts more tokens for the same text. The real saving is smaller than a third.

The crawler split, and the llms.txt data

Two GEO items. The first is the one to act on. Training crawlers (GPTBot, ClaudeBot, CCBot) feed model training. Retrieval crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot) fetch a page so it can be cited in an answer. Blocking one does nothing about the other, and the difference is the whole game: an audit of 95 well-known domains on September 1 found 17% still block PerplexityBot — precisely the fetch that would have quoted them.

Two operational notes. Cloudflare began changing its default configuration to block AI bots through September 2026, so a correct robots.txt is no longer sufficient on its own — check what your CDN and WAF actually allow. And ChatGPT Search retrieves from the Bing index, which makes Bing Webmaster Tools as load-bearing for AI visibility as Search Console is for organic.

The second item contradicts a lot of GEO advice, including advice this site used to repeat. SE Ranking analysed about 300,000 domains and found llms.txt on 10.13% of them, with no correlation to citation frequency — removing the factor from their gradient-boosting model made it more accurate. Ahrefs looked at 137,210 domains and found the overwhelming majority of published llms.txt files received no traffic at all. Google Search Central's position is that the file will neither help nor hurt.

The argument that stayed with me is John Mueller's: a self-reported file cannot help a model tell two competing sites apart, because every site would simply claim to be the best match. A signal anyone can assert about themselves stops working as a signal once enough sites assert it.

This site ships an llms.txt and allows every AI crawler, training ones included. On this evidence that is defensible — blocking GPTBot and ClaudeBot would cost no citations, and keeping Google-Extended allowed preserves Gemini grounding. But it should be a decision rather than an accident, and it is a decision we are open to changing. Tell us what you would do with the training crawlers and we will follow the argument, not the convention.

On this site

Cut your LLM API cost — the longer version of the caching arithmetic above.
GEO content that gets cited — what actually makes an answer engine quote a page.