OpenAI and Anthropic cut their flagship LLM API prices by up to 80% between June and September 2026, according to an analysis by Max Woolf published on his blog minimaxir.com. The cuts touched both frontier and budget-tier models, and Woolf's timeline shows they weren't isolated discounts but a running back-and-forth between the two labs over four months.
The clearest case is OpenAI's GPT-5.6 Luna. Launched June 26, 2026 at $1 input / $6 output / $0.10 cache-read per million tokens, OpenAI cut it by 80% on July 30 to $0.20 / $1.20 / $0.02, a move Woolf's post flags with "(???)" as unexplained at the time. He writes that he suspects "the 80% GPT-5.6 Luna price cut in July was meant to specifically compete with these smaller open-weight models," pointing to DeepSeek V4 Flash as the competitive trigger. Anthropic's parallel move was Claude Sonnet 5's launch pricing of $2 / $10 / $0.20 per million tokens, announced June 30 as temporary and made permanent on August 10.
The mechanics matter most on the cache-read line, not input/output pricing. Woolf's post explains that for coding agents, cache-read tokens — the stored context replayed every turn — make up "96% - 99%" of total token expenditure, with cache hit rates running "90 - 100%." That's why Claude Fable 5.1's September 1 pricing, which cut cache-read cost by 75% relative to Fable 5 while improving benchmarks across the board, "correlates to a 70-75% reduction in the cost of using the model" even though headline input/output prices stayed near $10/$50. The same logic applies to GPT-6.1 Sol, announced September 29 after the initial GPT-6 Sol release proved what Woolf calls "a dud performancewise": OpenAI halved its cache price to $2 / $10 / $0.10, making the GPT-6 Sol family "cumulatively 75% cheaper" than its predecessor.
On hardware and architecture, the sources name no GPUs or serving frameworks — this is a pricing analysis, not an infrastructure teardown. What it does carry is a trend line: data cited from Vercel's AI model gateway shows cost-per-token falling to 43.3% of its April 2026 level, even as Woolf notes aggregate spend is rising because "the models are getting better faster than they're getting cheaper." Anthropic's own framing of Opus 5.5, announced September 22 at $4 / $20 / $0.20 (down from Opus 5's $5/$25/$0.50), was explicit: "Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that," a combination of input/output cuts and cache reductions Woolf totals at "40-60%" lower cost for a model he says is better than Fable 5.1.
The competitive pressure cited is open-weight inference. Woolf counts 29 distinct providers hosting DeepSeek V4.1 Flash on OpenRouter, up from 13 for the original V4 Flash in May, with Cloudflare and DeepInfra priced "within 10% of DeepSeek's cost" on effective input cost. He also flags Moonshot AI's incoming Kimi K3.1 Pro, benchmarked by Artificial Analysis at 55 against Opus 5.5's 58, priced at half Opus 5.5's input/output rate with matching cache-read cost — weights due November 4, after which Woolf expects most of Kimi K3's 20-plus existing providers to pick it up at a comparable price.
What's unresolved is margin. Woolf cites unnamed reports putting inference profit-per-token margins "as healthy as 80% with caveats" before the cuts, but neither lab has disclosed what remains after them, and both OpenAI and Anthropic are, per the post, approaching IPOs — OpenAI in 2027, Anthropic "within a couple months" — a timing Woolf calls counterintuitive for companies about to court public-market scrutiny of "massive negative profits." He also notes OpenAI simultaneously cut the Codex $200/month tier's usage allowance by 50%, suggesting subscription value, not base API price, may be the first lever pulled if margins tighten. Woolf's own usage changed accordingly: after GPT-6.1 Sol cut his effective cost by 75%, he dropped his Codex quota to one-fifth, calling the net effect "roughly neutral and the obvious choice."
The takeaway for anyone budgeting a RAG or agent pipeline against GPT or Claude endpoints: don't price contracts off headline input/output rates — cache-read pricing is where 96-99% of agentic token cost actually sits, and that's the line that's moved most, and fastest, in this cycle.