If you budgeted for AI this year, your budget is already out of date. Between late June and late September 2026, OpenAI and Anthropic cut the prices of their flagship models repeatedly, with reductions reaching 80% on some models. The cheapest way to run a capable AI model today costs a fraction of what it cost four months ago — and the labs did not make those cuts out of generosity. They were forced into them by open-weight competitors, undercutting inference providers, and a market where switching models is almost free.

For a business, that is good news with a catch: the same collapse that slashes your per-token bill makes it easier to spend more overall, lock into the wrong vendor, or mistake a cheaper model for a better outcome. Here is what actually changed, why it changed, and what to do about it.

Four months that repriced AI

Data scientist Max Woolf, writing as minimaxir, published the fullest timeline of the cuts this week. All figures below are dollars per million tokens, in input / output / cache-read order — the three numbers that make up an API bill.

Two patterns stand out. First, the cuts accelerated: the biggest moves all landed in September, with both labs announcing flagship models on the same day, September 22. Second, the cuts came with better models, not worse ones: Opus 5.5 costs 40–60% less in total than Opus 5 while performing better, and GPT-6.1 Sol matches the much larger GPT-6 Astra at a fraction of the price. This was not a clearance sale on obsolete stock. It was a repricing of the frontier itself.

One more number frames the story: Vercel’s AI gateway data, cited by Woolf, puts aggregate cost per token at 43.3% of April 2026 levels — less than half the price in about six months.

The cut that mattered most: cache reads

Headlines quote the 80% figure, but the most consequential cut was quieter: the price of reading cached tokens. An AI coding agent re-reads the entire conversation history on every turn. That cached context accounts for 96–99% of total token spending in agentic coding, Woolf found, with cache hits on 90–100% of turns. So Anthropic’s 75% cache-read cut with Fable 5.1 translates into a 70–75% cut in the real cost of running the agent, even though input and output prices barely moved.

That is why the September 29 GPT-6.1 Sol announcement mattered: halving its cache price to $0.10 per million cached tokens matched Anthropic at 5% of input cost and made the GPT-6 Sol family cumulatively 75% cheaper than its June predecessor. If your team runs coding agents or any long-conversation workflow, your bill is dominated by a line item you probably never examined — now a quarter of what it was.

The practical lesson: stop comparing models on input price alone. For agentic work, cache-read pricing and your cache hit rate decide the bill — get all three prices from your vendor before you choose.

Why the labs had no choice: open weights

Woolf’s central argument: no company cuts its main revenue source by 80% without outside pressure. That pressure came from open-weight models — weights anyone can download and run — served by a growing crowd of inference providers competing on price.

The clearest example is DeepSeek. Flash V4 went viral in April 2026 at $0.14 / $0.28 / $0.0028 per million tokens, with cache reads at just 2% of input cost — far below the 5–10% the big labs charged. Its successor, Flash V4.1 1017, scores 45 on the Artificial Analysis intelligence index versus 24 for its predecessor, at a similar $0.15 / $0.60 / $0.003. Woolf suspects OpenAI’s July 80% Luna cut was aimed directly at these small open models.

Distribution did the rest. On OpenRouter, where any provider can list a hosted model, the number of distinct providers hosting DeepSeek’s Flash model grew from 13 in May to 29 for the newer variant, Woolf found. With switching costs near zero — a different endpoint string, a toggle in a coding harness — providers have little choice but to compete on price, and the big labs have to follow or lose the workload.

The next challenger is already lined up. Moonshot AI’s Kimi K3.1 Pro, announced days ago, scores 55 on the Artificial Analysis index against 58 for Claude Opus 5.5, at half the input and output price ($2 / $10 / $0.20). Its weights are due November 4, and with the previous Kimi K3 already hosted by more than 20 providers, Woolf expects most to offer Kimi K3.1 Pro at a similar price almost immediately. Every new open-weight release now resets the price floor for everyone else.

IBM’s “90% for 10%” claim — and the independent check

The enterprise version of this story arrived on IBM’s third-quarter earnings call. CEO Arvind Krishna told investors that open-weight models now deliver “anywhere from 25 percent to 10 percent of the token cost of the frontier models” while being “90 percent as effective,” The Register reported. His conclusion: most routine work — summarizing an email, examining a contract, answering a customer call — does not need the most expensive model, and a useful on-premises setup now fits in two or three server racks.

Krishna is, as commentators noted, talking his own book: IBM would like to sell the servers and accelerators those racks contain. But an independent check this week suggests his price claim is conservative while his performance claim is generous. Analysts at Tensorust pulled OpenRouter’s public pricing and Artificial Analysis scores for 215 models and found the top open model, MiMo V2.6 Pro, scoring 46.3 against 57.6 for Claude Opus 5.5 — about 80%, not 90% — but at $0.54 versus $8.00 per million blended tokens, or 6.8% of the price. Their summary: 80% of the capability for one-fifteenth of the bill.

For routine, well-specified work, the cheap model is now clearly good enough and the savings enormous. For hard agentic work — where a small capability gap means the job fails rather than degrades — the frontier model still earns its premium. The mistake is buying one tier for everything; the opportunity is matching the tier to the task.

What this means for your AI budget

• Re-benchmark now, not annually. If your model choices date from before September, they are priced on a market that no longer exists. Re-run key workflows against current models and prices; many teams will find a cheaper model now passes the same quality bar.
• Budget per task, not per token. Token prices fell, but usage is rising faster: Vercel’s gateway data shows aggregate spending still increasing because models improve faster than they cheapen. The right budget unit is cost per completed task — a resolved ticket, a shipped feature, a finished report — not cost per million tokens.
• Split workloads by tier. Route routine summarization, drafting, classification, and first-pass review to small or open-weight models. Reserve frontier models for tasks where failure is expensive: complex reasoning, high-stakes decisions, and long agentic chains.
• Negotiate on cache, not just input. For agent workloads, cache-read pricing and cache hit rates dominate the bill. Get all three prices in writing, and ask what happens to your cache when you switch providers.
• Keep your exit open. The price war rewards buyers who can switch: favor tools that let you change models with a setting, not a rewrite, and avoid workflows that only run on one vendor’s proprietary features.

A practical reset checklist

Work through this list before your next renewal:

1. Split your last three months of AI invoices into input, output, and cache-read spend. If cache reads dominate, your biggest saving is a model or provider change, not a usage cut.
2. Re-run your top five prompts or agent workflows on two cheaper models and score the outputs blind against your current model.
3. Price one open-weight option for each routine workload — hosted, not self-built, unless you have the engineering team to run it.
4. Check whether subscription allowances changed while prices fell: when GPT-6.1 Sol launched, OpenAI also cut the usage allowance on its $200-per-month Codex tier by 50%, Woolf noted. Headline prices and subscription value do not always move together.
5. Set a quarterly review date — in a market moving this fast, an annual review means overpaying for nine months.

For small teams starting out, falling prices lower the cost of experimentation. Free and low-cost tools are a sensible place to prototype before committing to a paid stack — our guides to free AI tools and the WebReOn Tools collection are built for that kind of low-risk trial, and our WebReOn Chat platform shows what a multi-model setup looks like in practice.

Three traps in a falling market

Trap one: spending more because each unit costs less. Cheaper tokens invite more ambitious agents, longer contexts, and more retries. Without a per-task budget, a 75% price cut can still produce a larger bill. Apollo’s chief economist Torsten Sløk noted this week that profit margins and revenue per worker look flat across the S&P 493 — the S&P 500 minus its seven largest tech companies — despite heavy AI investment. Cheaper AI is not automatically profitable AI; measure the workflow outcome, not the token count.

Trap two: mistaking a price cut for a quality guarantee. GPT-6 Sol itself was, in Woolf’s words, “a dud performancewise,” replaced within a week. Newer and cheaper is not always better for your specific task. Test before you migrate, especially on the tasks your customers see.

Trap three: assuming prices only fall. Woolf argues the “prices will snap back” theory is weak — with 29 providers on OpenRouter, no duopoly can dictate price — but both labs are reportedly approaching public listings, Anthropic’s rumored within months and OpenAI’s in 2027. Public markets reward margins, and today’s prices reflect a land-grab phase. Build your architecture so a future price rise is survivable: portable prompts, portable data, and no single-vendor dependency.

FAQ

Did AI really get 80% cheaper?

Specific models did: OpenAI cut GPT-5.6 Luna by 80% in July 2026, and the GPT-6 Sol family ended September cumulatively 75% cheaper than its June predecessor. Across the market, Vercel’s gateway data puts cost per token at 43.3% of April 2026 levels. Your own saving depends on your models and your cache-read share.

Are open-weight models really as good as frontier models?

Not quite, on independent measures. IBM’s CEO claims about 90% as effective; an OpenRouter and Artificial Analysis check puts the top open model at about 80% of the top closed model — but at under 7% of the blended price. For routine work that trade is excellent; for the hardest tasks, the frontier still leads.

Should a small business switch to open-weight models now?

For routine workloads, yes — through a hosted provider, not your own infrastructure. Keep frontier models for high-stakes or complex work and re-test quarterly. The cheapest option that passes your quality bar is the right one, and that bar moves every few months.

Will prices keep falling?

The forces pushing prices down — open-weight competition, many inference providers, near-zero switching costs — are still strengthening. But IPO pressure and rising demand could slow or partially reverse the trend, so plan for both directions.

The bottom line

The AI price war of 2026 made the best technology cheaper and better at the same time, and handed the advantage to buyers who stay flexible. The winners will not be the companies that picked the cheapest model once, but the ones that re-benchmark quarterly, split workloads by tier, watch cache pricing as closely as input pricing, and keep their architecture portable enough to follow the next price cut. Your AI budget from June is history. Build the next one for the market that exists today.

Sources and further reading

• Max Woolf (minimaxir): LLM prices cut by up to 80%
• The Register: IBM CEO on open-weight models on-prem
• Tensorust: IBM says 90%, the data says 80%
• ai0.news: October 26, 2026 AI digest
• DEV: LLM prices fell up to 80% — what to change
• LavX: OpenAI and Anthropic slash LLM prices up to 80%