Anthropic and OpenAI now charge the same per token in most tiers, so the price sheet decides nothing. The bill per task does. Tier by tier, Claude against OpenAI's closest model and the two best open models, and how to test them on your own work.
How to read the tables
Models and prices change every month, so treat this as a map, then measure on your own prompts. Every figure was checked on 22 October 2026 (sources at the end).
- Input and Output: US dollars per million tokens. Anthropic and OpenAI at their list price; open models at their maker's own endpoint on OpenRouter.
- Score: the Artificial Analysis Intelligence Index, an independent aggregate of reasoning, knowledge, maths and coding tests. Higher is better. Closed models are shown at high reasoning effort.
- Agentic: Terminal-Bench, the share of real tasks the model completes in a terminal, in %. The closest public measure of how it behaves as an agent.
- Cost to run: what Artificial Analysis paid to run its whole test suite on that model. It's the fairest single number for cost per intelligence, because it counts the tokens each model really uses, tokenizer and verbosity included. Lower is better.
- Open models: for each tier, the closest open model on raw performance and the best on performance for the price. When one model is both, the second row is the best alternative.
Top tier: Opus 5.5 and Fable 5.1
The hardest problems: architecture, deep analysis, long autonomous work where a wrong answer is expensive.
| Model | Input | Output | Score | Agentic | Cost to run |
|---|---|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | 53.6 | 56.6 | $2,172 |
| Claude Fable 5.1 | $10 | $50 | 51.2 | 52.0 | $5,242 |
| GPT-6 Astra | $10 | $50 | 50.9 | 54.0 | $2,925 |
| MiMo-V2.6-Pro (open) | $0.44 | $0.87 | 46.3 | 34.8 | $207 |
| GLM-5.3 (open) | $1.40 | $4.40 | 44.8 | 41.9 | $2,503 |
Opus 5.5 is the best model measured, and the cheapest of the three closed flagships to run. Anthropic positions Fable 5.1 above it and prices it like Astra, but on this index Fable scores lower than Opus and costs more than twice as much to run. Unless your own tests show Fable winning your task, Opus is the better Anthropic buy. At maximum effort, Opus reaches 57.6, and the gap to everything else widens.
The open models are a tier below here. MiMo-V2.6-Pro, from Xiaomi, MIT-licensed, is both the closest open model on raw score and by far the cheapest: seven points behind Opus for a tenth of the cost to run. GLM-5.3 is the strongest open model as an agent, but it costs as much to run as Opus. Both trail the closed models by 10 to 22 points on agentic work, which is where this tier earns its price.
Mid tier: Sonnet 5.5
Where most production features live: support answers, drafting, agents that call your tools.
| Model | Input | Output | Score | Agentic | Cost to run |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | $2 | $10 | 46.8 | 43.9 | $1,027 |
| GPT-6.1 Sol | $2 | $10 | 50.2 | 51.5 | $521 |
| MiMo-V2.6-Pro (open) | $0.44 | $0.87 | 46.3 | 34.8 | $207 |
| DeepSeek V4.1 Flash (open) | $0.30 | $1.20 | 45.1 | 33.8 | $445 |
This is where my verdict on the Anthropic API comes from. Same price per token, but Sol scores higher than Sonnet, works better as an agent, and costs half as much to run. MiMo-V2.6-Pro matches Sonnet's score for a fifth of the cost, and DeepSeek V4.1 Flash comes close for under half; both are weaker agents. Sonnet at maximum effort is excellent (56.0), at seven times the cost of its high setting.
Small tier: Haiku 5.5
The workhorse tier: classification, extraction, routing, summaries, at volume.
| Model | Input | Output | Score | Agentic | Cost to run |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | 37.8 | 21.7 | $94 |
| GPT-6 Luna | $0.10 | $0.50 | 32.9 | 4.5 | $48 |
| GLM-5.3-Flash (open) | $0.15 | $0.50 | 41.8 | 32.8 | $306 |
| MiMo-V2.6-Flash (open) | $0.14 | $0.28 | 37.9 | 22.7 | $110 |
Anthropic's best tier for value. Same price as Luna, clearly better, and far better as an agent. Luna costs half as much to run because it does less. GLM-5.3-Flash is the open model that beats Haiku outright, at three times the cost to run. MiMo-V2.6-Flash matches Haiku almost exactly, at about the same cost. Watch the long-prompt surcharge: above 100,000 tokens of prompt, Haiku costs five times as much.
The pattern
| Tier | Price sheet | Best result | Cheapest to run, close behind | Best open model |
|---|---|---|---|---|
| Top | Opus is cheapest per token | Opus 5.5 | Opus 5.5 | MiMo-V2.6-Pro, 7 points behind, a tenth of the cost |
| Mid | Level | GPT-6.1 Sol | GPT-6.1 Sol | MiMo-V2.6-Pro, level with Sonnet, a fifth of the cost |
| Small | Level | Haiku 5.5 (GLM-5.3-Flash if open) | GPT-6 Luna | MiMo-V2.6-Flash, level with Haiku |
So: the hardest reasoning, Opus. Small jobs, Haiku is a fair pick. Production features in the middle, test Sol and MiMo-V2.6-Pro first. And for anything that runs as an agent, the closed models still lead clearly.
How to test them yourself
A table is a map, not an answer. The model that wins a benchmark can lose on your prompts, and your prompts are what you pay for.
- One key for all of them. OpenRouter serves Claude, GPT and the open models through one API and one bill. Switching is one string.
- Replay real traffic. Send a week of real prompts to each candidate and count the cost of the answers that passed your checks, not the cost per token.
- For code, use an agent. OpenCode with OpenRouter behind it gives the same task in the same repository to any model, switched with one command. That's how I find out whether a cheaper model can do a client's job before anyone commits to it.
- Pin the provider in production. OpenRouter's cheapest listing for an open model is often a compressed (4-bit) version that scores lower than the maker's own endpoint. The prices above are the makers'. Once you've chosen, pin the provider you tested.
Traps
- Long prompts change the price. Haiku costs five times as much above 100,000 tokens of prompt, and OpenAI's models cost more above 272,000. Opus, Sonnet and Fable keep one price across their 1M-token window.
- Same text, different token counts. Anthropic says its newer tokenizer produces about 30% more tokens for the same text than its previous one. I found no rigorous measurement against OpenAI's. The cost-to-run column already includes the difference.
- Open doesn't always mean open. Several "Max", "Prime" and "Pro" variants on OpenRouter are API-only, with no downloadable weights. Every open model in these tables has its weights published on Hugging Face.
- Data location. None of these keep data in Australia by default. For that, Claude runs on Amazon Bedrock in Sydney and Melbourne, at 10% over global routing.
Sources
Checked on 22 October 2026:
- Artificial Analysis leaderboard: Intelligence Index v4.3.2, Terminal-Bench and cost to run. Closed models at high reasoning effort; open models as Artificial Analysis ran them.
- OpenRouter models API: prices, from each maker's own endpoint.
- Anthropic pricing and OpenAI pricing: list prices and long-prompt surcharges.
- Hugging Face model cards: open weights and licences (MiMo-V2.6-Pro and -Flash and DeepSeek V4.1 Flash under MIT; GLM-5.3 under its own licence, GLM-5.3-Flash under MIT).