OpenAI and Anthropic cut agent model prices up to 80% in 2026

Between 26 June and 29 September 2026, OpenAI and Anthropic cut the API prices of their main coding and agent models by as much as 80%, according to a pricing timeline published by former BuzzFeed data scientist Max Woolf on his blog on 25 October 2026. The cuts ran through GPT-5.6, GPT-6, GPT-6.1, Claude Sonnet 5, Claude Opus 5 and Claude Opus 5.5. The deepest reductions landed on cached input tokens, the part of a bill that dominates agentic coding sessions.

OpenAI and Anthropic cut the per-token price of their main coding and agent models by up to 80% between June and September 2026, with the steepest cuts hitting cached input tokens, which make up 96% to 99% of the cost of a typical agentic coding session.

What changed

Both companies cut prices repeatedly across ten announcements between June and September 2026, with the deepest cuts landing on cached tokens, and the pattern shows up in usage data too: Vercel's AI Gateway Production Index, published 15 October 2026, found September 2026 gateway spend ran about 13 times December 2025's, while the average price per token fell 47%.

• 26 June 2026: OpenAI announced GPT-5.6 Sol at $5 / $30 / $0.50 per million input, output and cache-read tokens (the same price as GPT-5.5), plus a cheaper GPT-5.6 Luna at $1 / $6 / $0.10.
• 30 June 2026: Anthropic launched Claude Sonnet 5 at $2 / $10 / $0.20, a price it said at the time was temporary.
• 24 July 2026: Anthropic shipped Claude Opus 5 at $5 / $25 / $0.50.
• 30 July 2026: OpenAI cut GPT-5.6 Luna's price 80%, to $0.20 / $1.20 / $0.02.
• 10 August 2026: Anthropic made the Sonnet 5 launch price permanent.
• 21 August 2026: OpenAI cut GPT-5.6 Sol about 20%, to $4 / $20 / $0.40, for three months.
• 1 September 2026: Anthropic released Claude Fable 5.1 at $10 / $50 / $0.25, cutting the cache-read price 75% from Fable 5's $1.
• 3 September 2026: OpenAI released GPT-6 Astra at $10 / $50 / $1, matching Fable 5.1's input and output price but not its cache-read cut.
• 22 September 2026: OpenAI released GPT-6 Sol at $2 / $10 / $0.20, half the already discounted Sol price, plus GPT-6 Luna at $0.10 / $0.50 / $0.01. The same day, Anthropic released Claude Opus 5.5 at $4 / $20 / $0.20. Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads: input and output prices fell 20%, and the cache-read price fell 60%, from $0.50 to $0.20 per million tokens.
• 29 September 2026: OpenAI released GPT-6.1 Sol, which halved its cached input price again to $0.10 per million tokens, a cumulative 75% cut for that model line since GPT-5.6 Sol.

Open-weight models, led by DeepSeek and Moonshot AI's Kimi, processed more than seven in ten gateway tokens in September 2026 and took close to one in four dollars of spend, up from one in twenty dollars in June, per Vercel. Claude Opus 5.5 alone took more than one in ten gateway dollars in its first two weeks, the largest spend share Vercel recorded for any model launched in September.

Why it matters for agent builders

Cache reads decide how much an agent actually costs to run. An agent harness resends its full conversation history on every turn, so the cache built from that history accounts for 96% to 99% of total token spend in agentic coding, according to Woolf's analysis, and that cache gets hit 90% to 100% of the time depending on the inference provider. Anthropic priced Opus 5.5 to match: it cut cache reads 60%, against a 20% cut to input and output. For a team running Claude Code or Codex agents at volume, the list price understates the savings. Woolf cut his own Codex subscription usage cap to a fifth after GPT-6.1 Sol shipped, because the same subscription dollar bought roughly four times the usage.

Open-weight models keep the pressure on OpenAI and Anthropic's prices. DeepSeek V4.1 Flash's cache-read price runs about 2% of its input cost, versus the 5% to 10% typical of OpenAI and Anthropic models, which lowers its effective price further still. On OpenRouter, 29 distinct providers now host the DeepSeek V4.1 Flash 1017 variant, up from 13 providers for the original DeepSeek V4 Flash in May.

What to watch

What to watch next is whether these price cuts survive the IPOs both companies are reportedly preparing, OpenAI in 2027 and Anthropic within a couple of months, according to reporting Woolf cites. That timing raises a question: why would a company preparing to sell shares cut its main revenue source by up to 80%? Woolf points to compute efficiency gains: Anthropic said directly that Opus 5.5 "requires less compute to serve than Opus 5." He also points to competitive pressure from the growing list of inference providers reselling open-weight models. Also worth tracking is whether subscription usage allowances get cut before the underlying API price does. OpenAI already cut the usage allowance on its $200 Codex tier by 50% the same week GPT-6.1 Sol shipped, citing "increasingly efficient models."

Key takeaways

• OpenAI and Anthropic cut per-token prices on their main coding and agent models by up to 80% between 26 June and 29 September 2026.
• Cache-read tokens, which make up 96% to 99% of a typical agentic coding session's cost, saw the deepest cuts: Claude Opus 5.5 cut cache reads 60% against a 20% cut to input and output.
• Vercel's AI Gateway data shows September 2026 gateway spend ran about 13 times December 2025's, while the average price per token fell 47%.
• Open-weight models, led by DeepSeek and Moonshot AI's Kimi, grew from one in twenty gateway dollars in June 2026 to nearly one in four in September 2026.
• OpenAI and Anthropic are both reportedly preparing IPOs, which puts their inference profit margins under new scrutiny even as list prices fall.

FAQ

Why did OpenAI and Anthropic cut their API prices in 2026?

The two companies moved within days of each other for most of 2026, which points to direct competition between them. Anthropic has also said its cuts follow lower serving costs: it noted that Opus 5.5 "requires less compute to serve than Opus 5." Growing competition from open-weight models such as DeepSeek V4.1 Flash and Moonshot AI's Kimi, which other inference providers can host at similar quality for less, adds pricing pressure neither company can ignore.

What is a cache read and why does it matter for AI agents?

A cache read is the discounted price a model provider charges for tokens it has already processed earlier in the same conversation and kept in storage for reuse. Because an agent harness resends the full conversation history on every step, cache reads make up 96% to 99% of total token spend in agentic coding, according to Woolf's analysis. A cut to the cache-read price, such as the 60% cut Anthropic applied to Claude Opus 5.5, lowers the real cost of running an agent by more than the headline input and output prices suggest.

How much cheaper are AI agent models now than in mid-2026?

It varies by model, but the cuts run as high as 80%. OpenAI cut GPT-5.6 Luna's list price 80% on 30 July 2026, and its GPT-6.1 Sol, released 29 September 2026, charges $0.10 per million cached input tokens, 95% less than its standard input price. Anthropic's Claude Opus 5.5, released the same day as GPT-6 Sol on 22 September 2026, costs 40% less to run than Opus 5 on typical workloads. Separately, Vercel's AI Gateway data shows the average price per token across all models fell 47% between December 2025 and September 2026.

Source: Max Woolf's Blog Discussion: Hacker News

Developers comparing which coding agents and models deliver the best price for the work they do can browse current options in AI Agents Listing's coding category.