OpenAI and Anthropic have cut the cost of their flagship models by as much as 80% in roughly four months. That’s according to a detailed timeline that’s been getting a lot of attention on Hacker News. The author tracked every pricing move from June to late October 2026, and the pattern is hard to miss. The two labs are answering each other’s cuts, sometimes on the same day.
What stands out is that quality didn’t drop. By the author’s account, these cheaper models match or beat the ones they replaced.
📉 The Timeline in Short
The price war happened in small steps, then all at once:
- July 30: OpenAI cut GPT-5.6 Luna by 80%, to $0.20 input / $1.20 output per million tokens. OpenAI’s own chart suggests Luna would have lost to DeepSeek V4 Pro on cost-per-task without the cut.
- August 10: Anthropic made Sonnet 5’s “temporary” launch price of $2 / $10 permanent.
- September 1: Anthropic’s Fable 5.1 kept its $10 / $50 list price but cut cache reads by 75%, from $1 to $0.25.
- September 22: OpenAI launched GPT-6 Sol at $2 / $10, half the price of the discounted GPT-5.6 Sol. That same day, Anthropic shipped Opus 5.5 at $4 / $20 / $0.20 and said it “requires less compute to serve than Opus 5, and its pricing reflects that.”
- September 29: GPT-6.1 Sol arrived with cache reads halved to $0.10. The author puts that model family at about 75% cheaper overall.
🧠 Why Cache Pricing Is the Real Story
Most coverage looks at input and output prices. The author argues that’s the wrong number to watch for agentic work.
In agentic coding, every turn resends the whole conversation, and that comes from the cache. The author estimates cache reads make up 96% to 99% of total token spend in those workflows, with cache hit rates of 90% to 100%. So Fable 5.1’s 75% cut on cache reads works out to a 70-75% drop in real-world cost, even though the headline price never changed.
That should change how teams read pricing announcements. A cut to cache pricing can matter more than a cut to output pricing.
⚔️ Why Now? Commoditization Meets Open Weights
The author’s main point is simple: “no one cuts prices by 80% unless there is some externality in play.” Two forces are pushing here.
- Switching costs are close to zero. Tools like OpenCode, Zed Agent, and OpenRouter let you swap models with a toggle or a new endpoint string. When models are hard to tell apart “outside of subjective vibes,” loyalty is thin and there’s not much of a moat.
- Open-weight models set the floor. DeepSeek V4 Flash launched in April at $0.14 / $0.28 with cache reads at just $0.0028, and it went viral. OpenAI didn’t even have a small model in that class at the time. Once a cheap open model is good enough, every closed lab has to justify its price against it.
Google, Meta, and others still count as substitutes too. That makes it harder for any single lab to hold its prices up.
📊 Cheaper Tokens, Bigger Bills
This is the counterintuitive part. Vercel’s AI gateway data, cited in the piece, shows cost-per-token has fallen to 43.3% of April 2026 levels. But total spending keeps rising, because models are “getting better faster than they’re getting cheaper.”
That’s a classic Jevons effect: make something cheaper and people use more of it. Lower prices don’t shrink AI budgets. They open up longer agent runs, bigger contexts, and use cases that weren’t worth it before.
One caveat. The author is a single practitioner writing from personal use and public announcements, and some quality judgments (like GPT-6 Sol being “a dud”) are subjective. The pricing data itself comes from the labs’ own announcements, though.
🎯 What Practitioners Should Do
- Re-benchmark every quarter. The cost-quality frontier moved several times in four months. Last quarter’s best pick may be overpriced now.
- Model your costs around cache reads. If you run agents, cache pricing and hit rates drive your bill more than anything else.
- Build for portability. Keep your stack model-agnostic so you can switch when the next cut lands.
- Don’t expect savings to stick. Plan for usage to grow into whatever you save.
The tit-for-tat between OpenAI and Anthropic shows no sign of slowing, and open-weight labs keep pushing the floor lower. For buyers, that’s mostly good news. For the labs, it raises a harder question: how do you build a durable business when your main product keeps losing value this fast? The full timeline and charts are in the original post.