Token prices from the two biggest US labs fell fast between June and October 2026. A dated timeline from developer Max Woolf, published on October 25 and picked up on Techmeme, shows cuts of up to 80% on individual models and about 75% for a cache-heavy coding agent moving from GPT-5.6 Sol to GPT-6.1 Sol. The explanation (competition, open-weight pressure, cheaper serving) is Woolf's analysis, not a statement from either lab.
We cross-checked his prices against our own launch coverage. They match, so this post lays out the timeline, shows the cache-read math that explains most of the savings, and separates what is documented from what is opinion.
TL;DR: the numbers that matter
| Question | Short answer |
|---|---|
| How big are the cuts? | Up to 80% on one model (GPT-5.6 Luna), about 75% effective on GPT-6.1 Sol for agents, 40-60% on Opus 5.5 versus Opus 5 (Woolf's math) |
| What changed most? | Cache-read prices, which dominate agent bills |
| Who started it? | Not clear. Woolf sees tit-for-tat between OpenAI and Anthropic, plus open-weight pressure |
| Will prices rise again? | Woolf argues no; he expects subscription allowances, not list prices, to be cut first |
| What should I do? | Re-run your cost per task on current models; do not budget from a June price sheet |
All prices below are input / output / cache read per million tokens, the same format Woolf uses. Check the live Anthropic pricing page and OpenAI API pricing page before you budget.
The timeline, month by month
June 26. OpenAI announced GPT-5.6 Sol at $5 / $30 / $0.50 (same as GPT-5.5) and the small GPT-5.6 Luna at $1 / $6 / $0.10. June 30. Anthropic launched Claude Sonnet 5 at a launch price of $2 / $10 / $0.20.
July 24. Claude Opus 5 arrived at $5 / $25 / $0.50, matching Sol on price. July 30. OpenAI cut Luna by 80% to $0.20 / $1.20 / $0.02. Woolf's reading of OpenAI's own chart: without that cut, Luna would have lost to DeepSeek V4 Pro on quality versus cost per task.
August. Anthropic made the temporary Sonnet 5 launch price permanent on August 10. On August 21 OpenAI cut GPT-5.6 Sol about 20% to $4 / $20 / $0.40 for three months.
September. On September 1 Anthropic shipped Fable 5.1 at $10 / $50 / $0.25. Cache reads fell 75% from the earlier $1, with benchmark gains over Fable 5. Our Fable 5.1 launch breakdown has the details. On September 3, OpenAI launched GPT-6 Astra at $10 / $50 / $1 (our Astra launch post).
September 22. The biggest day. OpenAI released GPT-6 Sol at $2 / $10 / $0.20 and GPT-6 Luna at $0.10 / $0.50 / $0.01. The same day Anthropic announced Opus 5.5 at $4 / $20 / $0.20, stating that Opus 5.5 "requires less compute to serve than Opus 5, and its pricing reflects that" (quoted by Woolf). See our Opus 5.5 launch coverage and the GPT-6 Sol and Luna pricing post.
September 29. OpenAI followed with GPT-6.1 Sol at $2 / $10 / $0.10. Woolf calls GPT-6 Sol "a dud performance-wise" and says 6.1 matches Astra in his own use. We covered it in GPT-6.1 Sol launch pricing and benchmarks.
Why cache reads are the real story
Sticker prices for input and output tokens grab headlines, but agent bills are dominated by something else. A coding agent re-sends the entire conversation every turn, and the provider serves most of it from cache. Woolf puts cache reads at 96-99% of total token spend in agentic coding, with cache hit rates around 90-100%.
If that is roughly right for your workload, the cache-read price is nearly your whole bill. Under that assumption:
| Move | Cache-read price change | Approx. effect on an agent bill |
|---|---|---|
| Fable 5 to Fable 5.1 | $1 to $0.25 (-75%) | About -70-75% |
| GPT-5.6 Sol to GPT-6.1 Sol | $0.50 to $0.10 (-80%) | About -75% cumulative with list cuts, per Woolf |
| Opus 5 to Opus 5.5 | $0.50 to $0.20 (-60%) | -40-60% in total, per Woolf |
This lines up with our earlier work on Claude Sonnet 5.5's cache-read cut to $0.10 and our Opus 5.5 cost-per-task analysis. The lesson repeats: compare models on cost per finished task, with realistic cache hit rates, not on the headline input price. Our guide to prompt caching in Claude Code sessions shows how to keep that hit rate high.
One caveat: Woolf's 96-99% figure is his estimate from agentic coding. A chat app with short contexts, or a batch job with little reuse, will see a different mix.
Why did they cut? Three explanations, ranked by evidence
Woolf offers three explanations. None is confirmed by the labs.
1. Low switching costs and open-weight pressure. Swapping models is a toggle in a harness such as OpenCode or Zed, or a new endpoint string on OpenRouter. Woolf points to DeepSeek V4 Flash, which launched in April at $0.14 / $0.28 / $0.0028 and went viral. Its 2% cache-read ratio, versus 5-10% at OpenAI and Anthropic, makes real cost lower than the list price suggests. The newer DeepSeek Flash V4.1 1017 scores 45 on the Artificial Analysis index versus 24 for its predecessor, at $0.15 / $0.60 / $0.003, per Woolf. Our DeepSeek V4.1 Flash 1017 release post covers the model itself.
He adds that DeepSeek V4.1 Flash has 29 distinct providers on OpenRouter, up from 13 for the original, and that some, such as Cloudflare and DeepInfra, price within 10% of DeepSeek's own effective rate. That is hard to sustain a premium against.
2. Competitor tit-for-tat. The September 22 same-day launches and OpenAI's matching of Anthropic's 5%-of-input cache-read ratio on September 29 fit this. Woolf says Kimi K3.1 Pro, at $2 / $10 / $0.20 and scoring 55 against Opus 5.5's 58 on the index he cites, adds pressure from the open-weight side. Weights are due by November 4 (our Kimi K3.1 Pro post). Vercel's gateway data, shown in Woolf's post, has cost per token at 43.3% of April levels, though aggregate spend keeps rising because models improve faster than they get cheaper.
3. Cheaper serving. Anthropic's own statement about Opus 5.5 is the only first-party hint. If newer models truly need less compute per token, cuts are partly just passing along efficiency.
What about the IPO angle?
Woolf notes that both labs are rumored to be heading to IPOs, Anthropic within months and OpenAI in 2027, and that leaked financials show large losses. Cutting your main revenue source looks strange in that setting, so he guesses inference margins are healthy, citing reports of up to 80% per-token margins "with caveats". Treat this as speculation. We have not seen audited inference margins from either company, and we did not verify the leaked figures.
"Won't prices snap back to the true cost?"
A common claim among AI skeptics is that today's prices are subsidized and will rise like ride-hailing fares did. Woolf's rebuttal is structural: unlike Uber and Lyft, OpenAI and Anthropic are not a duopoly, entry barriers to inference are lower, and many hosts can serve open models at similar quality. His view is that the "true cost of a token" is where price equals marginal cost, which keeps falling.
That does not make the claim safe to bet on. Subscription value is a different lever. The same day GPT-6.1 Sol launched, Woolf says OpenAI cut the Codex $200/month tier allowance by 50% "to reflect our increasingly efficient models". We covered the earlier round in the GPT-6 Astra Codex usage-limit cut and the Reddit reaction to GPT-6 Sol limits. If capital gets tight, Woolf expects plan allowances, not API list prices, to move first.
What this means for what you build or pay
- Re-benchmark quarterly. A June cost model is wrong. Rerun a fixed set of your own tasks and compute cost per completed task.
- Model cache behavior explicitly. Ask your provider for cache hit rates and make sure your harness preserves prefixes. At 90%+ hit rates, cache-read price is your bill.
- Keep a cheap fallback. Small models fell hardest: Luna went from $1/$6 to $0.20/$1.20 within a month, then to $0.10/$0.50 with GPT-6 Luna. See our small model comparison for routing options.
- Right-size your subscription. Woolf dropped his Codex plan to one-fifth of the max quota after getting about 4x more usage per dollar from the same model, an outcome he says OpenAI and Anthropic probably did not intend. Check whether you are paying for headroom you no longer need.
- Do not lock in long commitments on today's prices. Enterprise commitments are a moat both labs are building, per Woolf, so negotiate with the trend in mind.
What is verified and what is not
| Claim | Status |
|---|---|
| List prices for GPT-6.1 Sol, Opus 5.5, Kimi K3.1 Pro | Match our launch posts built from the labs' pages |
| Cache reads are 96-99% of agent spend | Woolf's estimate for agentic coding, not a lab figure |
| Vercel gateway cost per token at 43.3% of April | Shown in Woolf's post via Vercel; we did not re-pull it |
| Cuts are reactions to each other or to open models | Woolf's inference; labs have not said |
| Inference margins near 80% | Reported with caveats; not confirmed |
| Codex $200 tier allowance halved | Reported by Woolf; check OpenAI's current plan page |
The bottom line
The headline price drop is real and traceable to launch pages. The deeper takeaway is mechanical: for agent workloads, cache-read prices decide your bill, and both labs have moved those down hard. Why they did it is still opinion, with open-weight competition and lower serving cost the most defensible explanations. Plan for prices to keep drifting down, plan allowances to be less generous than they look, and measure cost per task rather than trusting any sticker.
Related reading
- GPT-6.1 Sol: launch, pricing and benchmarks
- Claude Opus 5.5: launch, benchmarks and pricing
- Claude Fable 5.1 and Mythos 5.1 launch
- Opus 5.5 task cost, cache and effort in Claude Code
- Kimi K3.1 Pro open weights and price
- DeepSeek V4.1 Flash 1017 open weights
- Claude Sonnet 5.5 cache-read cut
Prices and figures are as reported on October 25-26, 2026 and may change. Check each provider's live pricing page before you budget.