Haiku 5.5 cut 90% in October, but only if you change a model ID. Mistral Large 4's sale ended and Vertex retired Gemini 2.5 without asking, and the replacement OpenAI names for gpt-5-mini on December 11 costs 6.8x more.

If you read one paragraph

The biggest cut of the month is Claude Haiku 5.5 at a tenth of Haiku 4.5's price, and you only get it by changing a model string. Meanwhile Mistral Large 4 quietly doubled when its two-week sale ran out, Vertex retired Gemini 2.5 on October 20, and OpenAI has a December 11 shutdown list whose named replacement for gpt-5-mini costs 6.8 times as much. Check the last one first.

Everything that moved in October

Input / output per million tokens unless marked. Read October 25, 2026 from provider pricing, changelog and deprecation pages; the Kimi figures come from OpenRouter because Moonshot's own table renders empty. Perplexity's changelog gives the month but not the day. Nano Banana 2.1 also tripled input to $1.50, which matters for image editing.

We've covered most of these on the day. The two we hadn't: the Mistral Large 4 sale had no end date on October 10, and now it has one in the past. Mistral's changelog says "50% off for 2 weeks", and the model page has dropped the struck-through price. And By October 22 MiniMax had cancelled the end of its own M3 promotion. A launch discount that turns permanent is rare enough that we checked the page twice.

What doing nothing costs

Here's every change in October and the next ten weeks that reaches you whether or not you touch your code. We priced the same month on each: 500 million input tokens, 100 million output, no caching. Where a model shuts down, "after" is the replacement the provider names.

• gpt-5-mini to GPT-5.6 TerraDec 11$325 to $2,200 (6.8x)
• Gemini 2.5 Flash-Lite to 3.5 Flash-Lite (Vertex)Oct 20$90 to $400 (4.4x)
• gpt-5-nano to GPT-5.6 LunaDec 11$65 to $220 (3.4x)
• gpt-5 to GPT-5.6 SolDec 11$1,625 to $4,000 (2.5x)
• gpt-4.1-nano to GPT-5.6 LunaOct 23$90 to $220 (2.4x)
• o3 to GPT-5.6 SolDec 11$1,800 to $4,000 (2.2x)
• Mistral Large 4, sale ends~Oct 20$549 to $1,098 (2.0x)
• Gemini 4 Argon, intro price endsDec 1$2,000 to $4,000 (2.0x)
• Gemini 3.8 Flash, intro price endsJan 1$750 to $1,500 (2.0x)
• Claude Sonnet 4.5 to Sonnet 5.5Nov 30$3,000 to $2,000 (0.7x)
• gemini-3.5-flash rerouted to 3.6 FlashOct 8$1,650 to $750 (0.5x)

Bar length is the after/before multiple. gpt-5-mini: 500M at $0.25 plus 100M at $2.00 is $325; Terra at $2 / $12 is $2,200. Mistral: $0.68 / $2.09 is $549, $1.36 / $4.18 is $1,098. Vertex Flash-Lite uses 3.5 Flash-Lite, the cheapest Gemini replacement Vertex lists today; on October 11 it also listed 3.1 Flash-Lite at $0.25 / $1.50, which would make it $275. Haiku 5.5's prompts are assumed under 100K tokens.

Two lines go the right way. Google sent gemini-3.5-flash traffic to 3.6 Flash on October 8 and cut the bill by more than half without anyone asking. Sonnet 4.5's replacement is cheaper on the rate card too, though Anthropic's newer tokenizer counts more tokens for the same text, so the real saving is smaller than a third.

Everything else goes up, and most of it lands on December 11.

December 11: don't take OpenAI's suggestion literally

OpenAI shuts down gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro that day. Its deprecations page maps each to a GPT-5.6 model, which made sense when the list was written. Since then GPT-6.1 Sol and GPT-6 Luna have shipped, and they're cheaper than the 5.6 models they'd stand in for.

Same 500M in / 100M out month. List rates: gpt-5 $1.25 / $10, o3 $2 / $8, gpt-5-nano $0.05 / $0.40, GPT-5.6 Sol $4 / $20, Terra $2 / $12, 5.6 Luna $0.20 / $1.20, GPT-6.1 Sol $2 / $10, GPT-6 Luna $0.10 / $0.50.

The gpt-5-mini row is the one that stopped us. Follow the deprecation page and a $325 month becomes $2,200. GPT-6 Luna does the same job for $100, a third of what you pay today. It's not a like-for-like swap, but it scores 38 on Artificial Analysis at max effort, one point above GPT-5.6 Luna's 37. But a mini-tier workload that runs fine on gpt-5-mini is exactly the kind we'd test on GPT-6 Luna before paying for Terra.

For gpt-5 and o3, GPT-6.1 Sol at $2 / $10 costs half of what the named replacement does. There's a second reason to skip 5.6 Sol. Its $4 / $20 is a promotion OpenAI only promises "at least through November 21", and no price after that is published. Moving onto it three weeks after that date means signing up for a number nobody has printed.

The calendar to January 14

1. Oct 31, 16:00 UTC

   DeepSeek's free Search and Fetch calls through Harness, Claude Code and Codex end. $0.75 per 1,000 successful calls after. Our write-up.

2. Oct 31

   Moonshot closes the Kimi K2 models to new sign-ups. All K2 models retire by the end of December.

3. Nov 4

   Moonshot says Kimi K3.1 Pro weights ship by this date. If they do, third-party hosts can start undercutting $2 / $10. Our write-up.

4. Nov 19

   Vertex retires gemini-3.6-flash and names 3.8 Flash, at the same price, as its replacement.

5. Nov 21

   The earliest date GPT-5.6 Sol's $4 / $20 promotion can end. OpenAI hasn't said what follows. Our write-up.

6. Nov 30

   claude-sonnet-4-5-20250929 retires on the Claude API. Sonnet 5.5 at $2 / $10 is the named replacement.

7. Dec 1

   Gemini 4 Argon goes from $2 / $10 to $4 / $20. OpenAI also shuts down gpt-image-1-mini and gpt-image-1.5. Our write-up.

8. Dec 11

   gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro shut down.

9. Jan 1, 2027

   Gemini 3.8 Flash and 3.6 Flash leave their introductory price: $0.75 / $3.75 becomes $1.50 / $7.50. Our write-up.

10. Jan 14, 2027

    GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down, all pointed at GPT-6 Luna. Our write-up.

The Gemini 3.8 Flash date is the one with the most money behind it, because Flash carries far more traffic than Argon. We worked through that in August, when we listed every dated discount. Nothing since has moved it.

Seven weeks of grep

Grep your codebase for gpt-5-mini, gpt-5-nano, o3 and claude-sonnet-4-5. Those IDs stop answering within seven weeks, and the cheapest destination for most of them isn't the one on the deprecation page.

If you're still on Haiku 4.5, move. It's 90% off for prompts under 100K tokens, and Haiku 5.5 is still half price above that. The same goes for anyone on Kimi K3 or MiniMax M3 who hasn't looked at the point releases.

And if you budgeted Mistral Large 4 off its launch price, that line in your spreadsheet has been wrong since around October 20. Our pricing table now carries the list price. Run your own token mix through the calculator.

Sources

• OpenAI deprecations, OpenAI pricing and GPT-5.6 Sol - October 23 and December 11 shutdowns, replacements, Sol promotion wording
• Anthropic model deprecations and pricing - Sonnet 4.5 retirement, Haiku 5.5 rates
• Mistral changelog and Mistral Large 4 model page - two-week launch sale, current list price
• Gemini API pricing, Gemini deprecations and Vertex model versions - Argon and Flash price dates, Gemini 2.5 and 3.6 Flash retirements, Nano Banana 2.1
• DeepSeek Search guide, Kimi platform pricing, MiniMax pricing and Perplexity changelog - free-call cutoff, K2 retirement, M3 permanent discount, Decider price
• OpenRouter and Artificial Analysis - Kimi K3.1 Pro rates, GPT-6 Luna index score