Claude API pricing: what 20 models cost a coding agent
Every Claude model in the catalog that has been ranked in an agent category, priced three ways — list, cache read, and what the two come to at the token mix an agent actually sends. Recomputed from the source catalog on 2026-10-11.
List price is not the bill. A coding agent re-reads its context every step, so about 95.6% of the tokens it sends are cache reads — and how deep Claude discounts a cache read decides the bill more than the number printed in the row. That discount is a vendor policy, not a per-model one, and no published rate card puts it next to the other vendors’.
What Claude charges per million tokens
20 Claude models that have been ranked in an agent category of the Design Arena, read from OpenRouter’s public catalog on 2026-10-11. Three prices per row: what the row lists, what a cache read costs, and what the two come to at the token mix a coding agent actually sends.
Claude does not have one cache-read rate — it has 3. Across 20 rows the discount runs from 2.5% to 10.0% of that row’s own input price (2.5%, 5%, 10%). So a cheaper list price here can still be the dearer call once an agent starts caching, and no single discount figure describes this vendor. Repriced at a coding agent’s mix, Claude’s list input price overstates what an agent pays by a median 6.6× (range 6.6×–12.6×).
2 of these 20 rows cost more than the price in their own row once the prompt is long enough. The threshold reprices every token in the request, including the ones under it, so one token over the line roughly doubles the call. The thresholds are in the table.
10 rows are marked batch — the batch entries the catalog lists separately. They are kept apart on purpose: folding them in would read as if a normal call cost half of what it does.
| $ / 1M at agent mix | $ in / 1M | $ cache read / 1M | $ out / 1M | Model | Context | Long-context step | Best agents rank |
|---|---|---|---|---|---|---|---|
| $0.1508 | $1.00 | $0.10 | $5.00 | Claude Sonnet 5 batch |
1M | — | #5 htmlslides |
| $0.206 | $2.00 | $0.10 | $10.00 | Claude Opus 5.5 batch |
1M | — | #2 mobileapps |
| $0.2263 | $1.50 | $0.15 | $7.50 | Claude Sonnet 4.5 batch |
1M | $3.00 past 200k | #31 mobileapps |
| $0.2263 | $1.50 | $0.15 | $7.50 | Claude Sonnet 4.6 batch |
1M | — | #12 godotgamedev |
| $0.3017 | $2.00 | $0.20 | $10.00 | Claude Sonnet 5 | 1M | — | #5 htmlslides |
| $0.3771 | $2.50 | $0.25 | $12.50 | Claude Opus 4.5 batch |
200K | — | #17 androidnative |
| $0.3771 | $2.50 | $0.25 | $12.50 | Claude Opus 4.6 batch |
1M | — | #8 agenticgamedev |
| $0.3771 | $2.50 | $0.25 | $12.50 | Claude Opus 4.8 batch |
1M | — | #1 agenticslides |
| $0.3771 | $2.50 | $0.25 | $12.50 | Claude Opus 5 batch |
1M | — | #1 mobileapps |
| $0.3956 | $5.00 | $0.125 | $25.00 | Claude Fable 5.1 batch |
1M | — | #1 fullstack |
| $0.4121 | $4.00 | $0.20 | $20.00 | Claude Opus 5.5 | 1M | — | #2 mobileapps |
| $0.4525 | $3.00 | $0.30 | $15.00 | Claude Sonnet 4.5 | 1M | $6.00 past 200k | #31 mobileapps |
| $0.4525 | $3.00 | $0.30 | $15.00 | Claude Sonnet 4.6 | 1M | — | #12 godotgamedev |
| $0.7542 | $5.00 | $0.50 | $25.00 | Claude Fable 5 batch |
1M | — | #1 agenticgamedev |
| $0.7542 | $5.00 | $0.50 | $25.00 | Claude Opus 4.5 | 200K | — | #17 androidnative |
| $0.7542 | $5.00 | $0.50 | $25.00 | Claude Opus 4.6 | 1M | — | #8 agenticgamedev |
| $0.7542 | $5.00 | $0.50 | $25.00 | Claude Opus 4.8 | 1M | — | #1 agenticslides |
| $0.7542 | $5.00 | $0.50 | $25.00 | Claude Opus 5 | 1M | — | #1 mobileapps |
| $0.7911 | $10.00 | $0.25 | $50.00 | Claude Fable 5.1 | 1M | — | #1 fullstack |
| $1.5084 | $10.00 | $1.00 | $50.00 | Claude Fable 5 | 1M | — | #1 agenticgamedev |
Cheapest Claude row an agent can call normally is Claude Sonnet 5 at $0.3017 per million; the dearest is $1.5084, 5× more. Both numbers exclude the batch rows above. Both are computed, not quoted — the arithmetic and the weights are in the JSON.
Claude against the other vendors, on the same arithmetic
Same catalog, same day, same token mix. The column that decides an agent’s bill is not the list price — it is how deep that vendor discounts a cache read, because about 95.6% of what an agent sends is a cache read.
| Vendor | Rows | Cache read, % of its own input | List price overstates the agent bill by | Cheapest non-batch row at agent mix |
|---|---|---|---|---|
| xiaomi | 1 | 0.8% | 18.4× | $0.0237 |
| DeepSeek | 1 | 8.3% | 7.9× | $0.0264 |
| Claude | 20 | 2.5–10% | 6.6× | $0.3017 |
| Gemini | 12 | 10% | 6.6× | $0.0769 |
| OpenAI | 9 | 10–10.4% | 6.3× | $0.2042 |
| Llama | 3 | 12% | 6.0× | $0.2067 |
| Grok | 5 | 15–25% | 5.0× | $0.2494 |
| thinkingmachines | 1 | 17% | 4.7× | $0.215 |
| Qwen | 1 | 20% | 4.2× | $0.355 |
| GLM | 8 | 18.3–98.3% | 4.1× | $0.0646 |
| upstage | 1 | 20% | 4.1× | $0.0219 |
| MiniMax | 1 | 20% | 4.1× | $0.0731 |
| Kimi | 5 | 10–68.8% | 3.2× | $0.1711 |
The spread in that fourth column is the whole point: Claude at 6.6× against Kimi at 3.2×, across 20 and 5 rows. Two rows with the same list price, one from each, are not the same price to an agent — and no published rate card puts those two numbers next to each other.
All models, every vendor, one table · Put your own token counts in · JSON · CSV
What was written about Claude while these were measured
26 figures in these field notes come from a sentence that names Claude. These are quoted from the write-up, dated the day it went out — unlike the table above, they are not recomputed, so read each one as of its own date.
- 4 hours — “One developer pulled this off using Claude Opus 5, a fully automated multi-agent system, and a cycle that ran every 4 hours.” (2026-10-02) →
- 90% — “Higher accuracy than Claude Code — and a 90% drop in token usage.” (2026-10-01) →
- 80% — “The Claude Code team itself found that 80% of system prompts degrade performance.” (2026-10-01) →
- 5-hour — “The “5-hour limit” on AI coding agents like Claude Code is the primary bottleneck for independent developers, forcing mid-flow halts just as you hit a rhythm.” (2026-09-25) →
- 3 days — “One Claude Code user built an Android app for reading community content in spare moments by chatting and building in the same loop, and shipped 16 iterations in 3 days.” (2026-09-25) →
- 80% — “Claude Code’s 80% prompt reduction achieved identical performance metrics, proving that excessive constraints don’t improve outcomes.” (2026-09-17) →
- 80% — “Over-constraining prompts can hinder AI reasoning; for example, Claude Code’s team reduced their system prompt word count by 80% without any performance decline.” (2026-09-17) →
- 400 tokens — “For example, Anthropic’s frontend-design (approximately 400 tokens) uses a two-pass method for aesthetic guidance and has been installed over a million times.” (2026-09-15) →
- 20% — “Qwen’s cache read is 20% of its input where Anthropic’s is 10%, and at 96% cache reads that ratio is the invoice.” (2026-08-24) →
- 20% — “So this is Alibaba’s own policy against Anthropic’s own policy, 20% against 10%, with no reseller standing in between adding a markup I would have mistaken for a vendor decision.” (2026-08-24) →
- 10.00% — “Claude Sonnet 5 is resold through nine storefronts including AWS Bedrock, Azure and Google, and all nine bill a cache read at exactly 10.00% of their own input price.” (2026-08-24) →
- 10.00% — “Claude Sonnet 5 has nine, and all nine bill 10.00%.” (2026-08-24) →
- 10% — “Anthropic, Google and OpenAI each charge 10% of input for a cache read — flat, every model, and stable across sellers too.” (2026-08-24) →
- $355.00 — “Run the billion tokens through: Qwen3.7 Max bills $355.00, Claude Sonnet 5 bills $301.68.” (2026-08-24) →
- $2.00 — “Claude Sonnet 5 lists at $2.00.” (2026-08-24) →
- 80% — “For instance, Claude Code’s efficient programming capabilities, achieved by removing 80% of system prompts, which show these tools’ potential, allow independent developers to automate document processing, data analysis, and other tasks, thus benefiting businesses by improving efficiency.” (2026-08-22) →
- 80% — “When the Claude Code team decided to slash 80% of their system prompts, most developers expected the model to lose its edge in complex engineering tasks.” (2026-08-20) →
- $1.25 per million — “Meta priced Muse Spark 1.1 at $1.25 per million input and $4.25 per million output, roughly 75% and 83% below Anthropic’s Opus, and the tradeoff is visible in the benchmarks, since it leads on MCP Atlas and JobBench while trailing on SWE-Bench Pro and DeepSWE 1.1.” (2026-08-19) →
- 400 tokens — “Anthropic’s SKILL.md file is 400 tokens of aesthetic guidance and a two-pass working method, and it has been installed over 1.08 million times — a return on 400 tokens that argues taste-shaping is worth more than tool-building, at least in front-end generation.” (2026-08-10) →
- $19 billion — “Anthropic’s $19 billion data center lease locks in five to eight years of compute, which reframes the competition as a contest over power and floor space rather than architecture.” (2026-08-10) →
- $1.25 per million input tokens — “Meta’s Muse Spark 1.1 agent model API is priced at $1.25 per million input tokens and $4.25 per million output tokens — roughly 75% and 83% below Anthropic Opus on input and output respectively — and the target is explicit: become the cheap entry point for agent workflows and pull developers into the ecosystem.” (2026-08-10) →
- 80% — “Claude Code’s team discovered that removing 80% of system prompts actually improved programming performance, revealing how excessive model constraints can hinder rather than help AI effectiveness.” (2026-08-07) →
- 400-token — “Anthropic’s 400-token SKILL.md file, through its “two-pass workflow” and specific aesthetic guidance, has achieved over 1 million installations, proving that aesthetic direction is more useful than mere tool innovation.” (2026-08-07) →
- 400 tokens — “Anthropic’s SKILL.md file runs about 400 tokens, uses a two-pass approach with specific aesthetic guidance, and has passed 1.08 million installations.” (2026-08-05) →
- $19 billion — “Anthropic signed a data center lease reported at $19 billion, which is the kind of commitment that only makes sense if compute, not model architecture, is the constraint that decides who is still standing in five years.” (2026-08-05) →
- $19 billion — “Anthropic is not signing a $19 billion lease in order to cut prices in the next 12 months.” (2026-08-05) →
- 1.6 Billion Free Tokens Is a Compression Ratio, Not a Strategy · reply box · telegra.ph
- AI Model Costs: Beyond Per-Token Pricing · reply box · telegra.ph
- Chinese Models Are Not 2x Cheaper Once Your Agent Starts Caching · reply box · telegra.ph
- Claude Code Can Model a Flange but Not a Freeform Surface: A Six-Step Handoff Checklist for Non-Code Agent Output · reply box · telegra.ph
- Claude Code and Codex for Office Automation · reply box · telegra.ph
- How Alibaba’s Open Code Review Slashed AI Code Review Costs by 90% — And What It Means for Independent Developers · reply box · telegra.ph
- How Chinese AI Agent Tools Leverage 1.6 Billion Free Tokens · reply box · telegra.ph
- How a Multi-Agent AI System Made $7,600 in 7 Days for Under $100 in API Costs · reply box · telegra.ph
- Stop Hitting the 5-Hour Limit: Routing Your IDE’s AI Requests to Local Models · reply box · telegra.ph
- The Hidden Costs of Over-Prompting in AI Coding: Lessons from Claude Code’s Optimization · reply box · telegra.ph
- The Token Cost War: Why Price per Million Tokens Now Decides the AI Market · reply box · telegra.ph
- Why Stripping 80% of System Prompts Actually Improved Claude Code’s Performance · reply box · telegra.ph
A Claude price that looks wrong? Say which row — the form already knows it is about Claude; you only have to say what the number is now. Every figure here is computed from a published catalog, so a wrong one is a bug, not an opinion.
Did this answer it? A star on the repository is the whole ask — the data is CC BY and does not require starring.