Claude Sonnet vs Opus: Opus costs 2x per token but only 10% more per task, and scores 2 points higher

Which is better, Claude Sonnet or Opus?

Claude Opus is better on 6 of the 10 tests in the index, and Sonnet on 3. The tenth, GDP.pdf, is a tie at 26%. Artificial Analysis ran the September 2026 releases of both at max effort, and we read its comparison page on 12 October 2026. The widest Opus lead is on AA-Omniscience, a test that scores a right answer up and a confident wrong one down: 46 to 32. All three of the tests that Sonnet wins are agent tests. It's ahead on Terminal-Bench 4.0 by 64% to 60%, on AutomationBench-AA by 72% to 70%, and on AA-Briefcase by 1,823 Elo to 1,807.

All of the scores, costs, times and speeds in the table come from Artificial Analysis, which ran its Intelligence Index v4.3.2 on all four of them at max effort. The token prices are from Anthropic's pricing docs and the plan columns from claude.com/pricing, both read the same day. Haiku's price holds for prompts up to 100,000 tokens. Past that, Anthropic charges $0.50 and $2.50.

Context won't decide it. Anthropic lists the same 1M token context window and the same 128K token reply limit for all four, as of October 2026.

How many Sonnet answers does one Opus answer cost?

It costs two at the list price, and 1.1 when you let the model decide how long it works. Anthropic's price list makes Opus exactly twice Sonnet on every line, as of October 2026. You pay $4 against $2 for each million tokens you send, and $20 against $10 for each million it writes. A cache read is $0.20 on Opus and $0.10 on Sonnet, and the Batch API takes half off all of them. So a chat answer of 1,000 tokens in and 500 out costs $0.014 on Opus and $0.007 on Sonnet.

Agent work flips that. In the Artificial Analysis run, Sonnet wrote 197,000 output tokens per task, and 147,000 of them were thinking. Opus wrote 119,000. Sonnet spends its half price on 66% more output tokens. Output alone still leaves Opus 21% dearer, at $2.38 a task against $1.97. The rest of each bill, $3.60 for Opus and $3.49 for Sonnet, pays for input and cache, where Sonnet takes in about twice the tokens at half the price. So an Opus task costs $5.98 and a Sonnet task $5.46. That's 10% apart. The whole index bill shows a wider gap, $8,708 for Opus against $7,260 for Sonnet, or 20% more. Both are a long way under the 2x that the token price suggests. Opus also finished sooner, at 776 seconds a task to the 906 of Sonnet, even though Sonnet streams 135 tokens a second to 96.

We worked out the last two columns from the prices and scores above. Points per dollar put Sonnet 6% ahead of Opus, and Haiku about twenty times ahead of both. Fable comes last. So pay the token gap only when you fix the answer length, as in a support reply, a summary or a label. When an agent runs at high effort and picks its own length, Opus costs about what Sonnet does and it's done first. The cheapest way to use Claude Opus covers batch and cache discounts.

Token counts don't carry across generations, either. Anthropic says the tokenizer in its current models makes about 30% more tokens for the same text than the one older Claude models used. So a bill from the last Sonnet on the old tokenizer, at $3 and $15, doesn't map one to one onto today's $2 and $10.

Is Claude Sonnet or Opus better for coding?

Sonnet is better for a coding agent that works in a terminal, and Opus for code written from a spec. On Terminal-Bench 4.0, Sonnet scores 64% to 60% for Opus. On SciCode, which asks for scientific code from a written description, Opus leads 67% to 61%. Those two tests are the coding fifth of the Intelligence Index, read 12 October 2026.

That terminal score is new. A generation earlier, Sonnet 5 scored 14% on Terminal-Bench 4.0 and Opus 5 scored 49%. So in one release Sonnet went from under a third of what Opus could do in a terminal to a little ahead of it. Haiku scores 33% there. Fable scores 52%, which puts the priciest Claude third of the four at this kind of work.

Use Sonnet for the long coding sessions where a token bill adds up. Switch to Opus when the code rests on hard reasoning, since it also leads Humanity's Last Exam by 61% to 55%.

Is Claude Sonnet as good as Opus now?

Nearly, on the index. Sonnet 5 trailed Opus 5 by 13 points, 38 to 51, in the Artificial Analysis run we read on 12 October 2026. Anthropic shipped the next Sonnet in September 2026, the same month as the next Opus. It scores 56. That's 5 points above Opus 5 and 2 below the current Opus.

Opus still has a real lead in two of the tests. It scores 46 to 32 on AA-Omniscience and 61% to 55% on Humanity's Last Exam. If your question has one right answer and you can't easily check it yourself, those two gaps will matter to you more than the 2 point headline does.

Anthropic's own docs lean both ways. Its models overview, read 12 October 2026, tells developers who aren't sure to "start with Claude Opus 5.5 for most workloads." Its pricing page still names Sonnet for most production work and Opus for the hardest reasoning. See the smartest AI model for where both sit against GPT and Gemini.

Where do Claude Haiku and Fable fit against Sonnet and Opus?

Haiku is the cheap, fast tier, and Fable is the dear one that scores below both on the index. Anthropic released its newest Haiku in October 2026 at $0.10 and $0.50 per million tokens for prompts up to 100,000 tokens. That's a twentieth of the Sonnet price. It scores 43 on the index, it streams 235 tokens a second, and it costs $0.21 for each task. Anthropic says it's for the high volume jobs, such as classification, extraction and routing, where you need an answer fast.

Fable costs $10 and $50 per million tokens. That's 2.5 times Opus and 5 times Sonnet. At max effort it scores 53, behind both Sonnet and Opus. It writes fewer tokens than any of the other three, at 78,000 a task, and it still costs the most of them per task, at $7.63.

Anthropic says to use Fable for demanding reasoning and long agent runs, or when Opus at higher effort still falls short in your own tests. On the October 2026 index, there's no general case for paying 2.5 times what you pay for Opus. Your own tests may say otherwise.

Which Claude plans include Sonnet, Opus and Fable?

Claude Free includes Sonnet and Haiku, and Opus starts on Pro. That's from Anthropic's plan table, read 12 October 2026. Pro costs $20 a month, or $17 a month when you pay $200 for the year. Max starts at $100 a month for 5 times the usage of Pro, with a 20 times tier above it.

Fable works differently on each paid plan. Anthropic's Fable help article says Pro users pay for it with usage credits, outside the plan limits. If you're on Max, you can spend up to half of your weekly limit on Fable, and it won't cost you anything extra. Claude Pro vs Claude Max covers how those limits work.

Whizi Pro carries Opus and Sonnet on one $29.99 plan with 2,500 credits a month. As of October 2026, Opus costs 20 credits a message and Sonnet costs 10, so the plan buys 125 Opus or 250 Sonnet messages. That's the same 2 to 1 ratio as Anthropic's token prices, and side by side model chats send one prompt to both, billed at both rates. Whizi loses on Fable. It sits on Powerhouse at $49.99 and 100 credits a message, which comes to 100 Fable messages a month, while Claude Max puts it inside the plan limits. Whizi has no developer API either. The Claude in Whizi docs list every row.

Should you use Claude Sonnet or Opus?

Use Sonnet when you pay per token for answers of a set length, and Opus when an agent thinks at max effort or the facts have to be right. Each of the calls in the table below has the number behind it, read 12 October 2026.