Claude Haiku vs Sonnet vs Opus vs Fable: Models Compared

Which Claude model to start with depends on where you use it. In the Claude apps (claude.ai, desktop, and mobile), start with Sonnet 5.5: Anthropic's Claude Academy tutorial calls it the versatile default and tells you to start there if you're not sure. In the Claude API, Anthropic's model overview says to start most workloads on Opus 5.5 (claude-opus-5-5). In both places, Haiku 5.5 is the tier for narrow, high-volume jobs such as classification, extraction, routing, and subagent work, and Fable 5.1 is the one to test only when Opus 5.5 at a higher effort setting still falls short. That is the lineup as of October 26, 2026. Haiku 4.5, Sonnet 5, Opus 5, and Fable 5 are now legacy models.

The price gap between tiers is real, but it is smaller per task than per token. On prompts up to 100,000 tokens, Opus 5.5 costs 40 times as much per token as Haiku 5.5 and twice as much as Sonnet 5.5. At each model's default effort, an average task in Artificial Analysis's Intelligence Index costs $0.05 on Haiku, $0.88 on Sonnet, $1.34 on Opus, and $3.91 on Fable, and Fable is never the cheapest way to reach any score on that index. In a 9-task test run on October 26, 2026, Haiku, Sonnet, and Opus all passed every check; Opus cost about 31 times as much as Haiku and 3.1 times as much as Sonnet, and Sonnet finished first on 8 of the 9 tasks.

If a cheaper tier passes your own checks, stay on it. If it fails, raise its effort setting before you change models, because Anthropic's model guide says tuning effort is often a better lever than switching. Move up a tier when the current one still fails at a higher effort, and move down when the cheaper one passes the same prompts.

What separates Haiku, Sonnet, Opus, and Fable: price, effort, plan

The four current models share more than their names suggest. According to Anthropic's Claude models overview, all four have a 1M-token context window, 128K maximum synchronous output, a reliable knowledge cutoff of June 2026, text and image input, tool use, and vision. What differs is price, default effort, whether thinking can be turned off, Anthropic's relative speed label, and which Claude plans include them.

Prices are Anthropic's first-party API list prices per million tokens (MTok) on its Claude API pricing page, with global routing. Plan access and limit use come from the Claude Academy tutorial on choosing a Claude model, which says Free includes Haiku and Sonnet while Pro and Max add Opus, Fable, and more headroom. For what each paid plan's limits look like in practice, see Claude Pro vs Max vs Ultra: Usage Limits and When Max Is Worth It.

The chart shows the August 2026 lineup; the table and text above are current as of October 26, 2026.

The speed labels are Anthropic's relative ranking, not a measurement. The overview's own footnote says actual latency depends on prompt length, output length, and effort, and in the October 26 test described below, Sonnet 5.5 finished ahead of Haiku 5.5 on 8 of 9 tasks because Haiku wrote far more tokens. Anthropic commits to keeping all four available on the platforms it operates until at least September 2027: Fable 5.1 until September 1, Opus 5.5 until September 22, Sonnet 5.5 until September 28, and Haiku 5.5 until October 7, 2027. Claude Mythos 5.1, which has Fable's capabilities, is open only to organizations verified through Anthropic's verification programs, so it isn't a self-serve choice alongside these four.

Why the Claude apps say Sonnet and the API says Opus 5.5

Anthropic gives two different starting points, and both are current. The Academy tutorial for app users says to start with Sonnet 5.5 and reserve Opus for "problems where you've tested with Sonnet and it struggled." The API models overview says that if you're unsure, start with Opus 5.5 for most workloads. The difference comes from what each audience spends.

In the apps you pay a flat plan price, so the scarce resource is your usage limit. Anthropic ranks limit use as lightest for Haiku, moderate for Sonnet, heavy for Opus, and heaviest for Fable, and it publishes no exact ratio. Starting on Sonnet keeps most work at a moderate cost to your limit, and Opus is only available from Pro upward anyway. In the API you pay per token, and Anthropic's guide to choosing a Claude model is written for teams building agentic coding and knowledge work: implement on Opus 5.5, then lower effort or move to a cheaper model once your evaluations show the work still passes. The same guide offers an efficiency-first path for prototypes, tight latency, cost-sensitive deployments, and high-volume straightforward tasks: start on Haiku 5.5 and upgrade only for specific capability gaps.

So pick the starting point that matches your situation rather than treating one as wrong. A chat user drafting, coding, and analyzing on a Pro plan should start on Sonnet. A developer building a coding agent should start on Opus 5.5 at its default medium effort and work down. A pipeline that labels a million support tickets should start on Haiku 5.5 and work up.

Claude API price per million tokens, and what one request costs

Base prices are in the table above; caching and Batch change the bill more than most people expect. The rows below are Anthropic's list prices per million tokens as of October 26, 2026.

Two rows need care. Haiku 5.5 has a second price band: a prompt over 100,000 tokens costs $0.50 input and $2.50 output, five times the short-prompt rate, and the pricing page says the threshold counts all input tokens, cache reads and writes included. Sonnet 5.5's cache read dropped from $0.20 to $0.10 on October 7, 2026, which Claude Sonnet 5.5 Cache Read Price Cut to $0.10: What You Save works through. Requests pinned to the US with inference_geo cost 1.1 times these prices.

To see what that means for one request, multiply each token count by its price and divide by 1,000,000. These four request shapes were recomputed at list price on October 26, 2026, with no Batch discount and the same token counts on every model:

Request C shows where Haiku's price edge shrinks. In B, Sonnet costs about 16 times as much as Haiku ($0.039 against $0.0024); in C, the 150K-token prompt pushes Haiku into its higher band, and Sonnet costs only 3 times as much ($0.045 against $0.015). Long agent loops reach that band sooner than you might expect, since the cached conversation counts toward the threshold. The table also assumes equal token counts, which real tasks don't produce; the next two sections show how much that matters. The first cache write in B and C is left out.

Moving up a tier at the same quality: Fable is never the cheapest

Per-token ratios tell you little about what a better answer costs, because each model can run at five effort levels and uses a different number of tokens per task. Artificial Analysis runs every model at every effort level (low, medium, high, xhigh, max) through its Intelligence Index, an average of 10 evaluations, and reports the cost and time per task. Recomputed on October 26, 2026 from its Sonnet 5.5 vs Opus 5.5, Haiku 5.5 vs Sonnet 5.5, and Opus 5.5 vs Fable 5.1 comparison pages, the cheapest of the 20 configurations that reaches each score is:

Fable 5.1 never appears. At its default high effort it scores 51, the same as Opus 5.5 at medium, but costs $3.91 per task, 2.92 times as much, and takes 414 seconds, 2.09 times as long. Its ceiling at max is 53, below Opus 5.5's 58 and Sonnet 5.5's 56; Haiku 5.5 tops out at 43. The steps between defaults are uneven too: Haiku to Sonnet adds 13 points for 17.6 times the cost, while Sonnet to Opus adds 4 points for 1.52 times the cost, partly because Opus defaults to medium and Sonnet to high.

The same data settles a few questions people ask about mixing tiers and effort. Haiku 5.5 at high scores 38 for $0.08 per task, beating Sonnet 5.5 at low (36 for $0.35) on both counts. Opus 5.5 at low scores 42 for $0.55, about what Sonnet at medium (41 for $0.48) and Haiku at xhigh (41 for $0.12) reach for less. Above a score of about 51, raising Sonnet's effort costs more than switching: Sonnet at xhigh scores 52 for $2.01, while Opus at high scores 54 for $1.82.

Read these as one benchmark's task mix, not yours. Artificial Analysis labels the Sonnet, Opus, and Fable runs "with fallback" without defining the term, prices Haiku only at its short-prompt band, and re-measures over time. An index average can also hide the long-horizon work Anthropic sells Fable on, which is why Claude Fable 5.1 vs Opus 5.5: Which Should You Use? covers when Fable's premium is justified.

Does Opus use more tokens than Sonnet? A 9-task test on October 26

Yes, in this test Opus 5.5 wrote about twice as many output tokens as Sonnet 5.5, even though Opus ran at its lower default effort. On October 26, 2026, Haiku 5.5, Sonnet 5.5, and Opus 5.5 each received the same 9 tasks at their API default effort (Haiku medium, Sonnet high, Opus medium), with no system prompt, no tools, no caching, and one sample per task. The tasks were a word problem, five Python functions graded by hidden test suites (8 to 25 cases each), summarizing a 6,723-word article into exactly 5 bullets of 25 words or fewer, a combinatorics count, and fixing a billing module with 5 planted bugs and 2 missing functions against 36 hidden tests. The pass checks were fixed before any answer was read.

Every model passed every check, so these tasks don't show which model is smarter; they show what the same passing work cost and how long it took. Each model read the same 18,003 input tokens, so the differences come from output. Haiku wrote 2.74 times as many output tokens as Sonnet and Opus wrote 1.95 times as many, which is why the per-task gaps differ from the list ratios: Sonnet cost about 10 times as much as Haiku instead of 20, Opus about 31 times as much instead of 40, and Opus 3.11 times Sonnet instead of 2. On the long-document task, where input dominates, the ratios moved back toward the input list prices.

Speed went against the labels. Sonnet finished first on 8 of the 9 tasks, Haiku was fastest only on one, and Opus was slowest on all 9. Haiku generates tokens faster (Artificial Analysis measures 173–234 output tokens per second for Haiku and 105–134 for Sonnet), but writing almost three times as many tokens cancels that. Times include the gateway and streaming, so treat them as relative.

The calls went through LaoZhang's Anthropic-compatible /v1/messages endpoint, which lists the same per-token prices as Anthropic for these four models, and costs are the returned token counts multiplied by those list prices; the whole run cost $0.37. What wasn't tested: Fable 5.1, Anthropic's own endpoint, effort overrides, prompt caching, the Claude apps, and repeat samples. For the apps, Anthropic only says Opus uses your limit "heavy" and Sonnet "moderate," which points the same way.

Is Haiku better than Sonnet, and is Sonnet as good as Opus?

No on both counts for hard work, but often close enough for narrow work. On the Intelligence Index, Haiku 5.5 at its best setting (43) stays below Sonnet 5.5 at its default (47), and Sonnet's default (47) is below Opus 5.5's default (51). Anthropic's Haiku 5.5 launch post says Sonnet 5.5 and Opus 5.5 "remain better choices for complex agentic coding tasks" and positions Haiku for narrowly scoped work such as compaction, summarization, and subagent tasks. On well-specified tasks like the nine above, though, all three passed, so the cheaper model only loses when the task is hard enough to separate them.

Open-ended work tends to show the gaps that a pass/fail test hides. MindStudio's October 9, 2026 write-up of one creator's same-prompt projects found that Haiku 5.5 built a working game, Sonnet 5.5 produced better graphics at roughly five times the cost, and Opus 5.5 had the best visual fidelity; on a forensic-accounting task all three found the same 8 of 8 planted frauds. That is a single unrepeated test, but it matches the pattern: the more the output's quality is a matter of judgment, the more the higher tier earns its price.

For the pairs in more depth, Claude Haiku 5.5 vs Sonnet 5.5: When the 20x Price Gap Pays Off covers which Sonnet work can move to Haiku, and Haiku 5.5 vs Opus 5.5: 40x Cheaper per Token, 23–29x per Task covers a Haiku-first setup that escalates to Opus. If your Opus bill is the concern, Claude Opus 5.5 Pricing and Limit Reset: What Each One Changes breaks it down.

Raise effort before you switch to a bigger model

Effort is the cheaper first move because it keeps your prompts, tools, and request format unchanged. All four models accept low, medium, high, xhigh, and max through the Claude API effort parameter, and sending the default is the same as leaving it out. In the Claude apps the same control sits in the model picker; the Academy tutorial says the default suits most work, a lower setting gives quicker answers that use less of your limit, and the highest setting is for your hardest problems.

The signal to switch tiers is a failure that survives a higher effort. Anthropic's API guide puts it concretely: if your evaluations on Opus 5.5 at xhigh or max still fall short on demanding reasoning or long-horizon agent work, move to Fable 5.1. For Haiku 5.5, it says to use xhigh or max only where your evaluations show a gain, and to compare those runs with Sonnet 5.5 on quality, cost, and speed. The reverse signal is a cheaper tier, or a lower effort, that passes the same prompts. The matched-quality table above shows why both directions matter: Haiku at xhigh reaches Sonnet medium's score for a quarter of the cost, while Sonnet at xhigh costs more than Opus at high for a lower score.

You can also run two tiers together. Anthropic's model guide describes an executor that hands hard decisions to a stronger advisor, and an orchestrator that delegates bulk work to cheaper workers, so most tokens are billed at the lower rate. In Claude Code, Anthropic's cost documentation recommends setting model: haiku in the configuration of simple subagents. The built-in Explore subagent runs on your main conversation's model unless you define a subagent named Explore with model: haiku, or set CLAUDE_CODE_SUBAGENT_MODEL together with CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1; on Bedrock, Google Cloud, and Foundry, pin claude-haiku-5-5, because the haiku alias there resolves to an earlier version.

Haiku 4.5, Sonnet 5, Opus 5, and Fable 5: what they map to now

All four names from the August 2026 lineup are legacy models as of October 26, 2026: still available on the Claude API, but no longer the current tier. Anthropic's models overview lists them alongside Opus 4.8, 4.7, 4.6, and 4.5 and Sonnet 4.6 as legacy models. Each has a direct replacement at the same tier.

Sonnet 5's introductory $2 / $10 rate became its standard price; the rise to $3 / $15 that was scheduled for September 1, 2026, did not happen, according to Anthropic's pricing page. If you're deciding between Haiku 4.5 and Sonnet 5 today, make the same decision between Haiku 5.5 and Sonnet 5.5 instead: Haiku 5.5 costs one-tenth of Haiku 4.5 per token on prompts up to 100K tokens but only half as much above that, so long-prompt workloads gain less from the move. The older Claude Opus 4 and Sonnet 4, which this page's address still names, were retired on June 15, 2026, on the platforms Anthropic operates, according to Anthropic's model deprecations page; Amazon Bedrock and Claude Models on Google Cloud (Google managed) set their own schedules and still list them.

What breaks when you change the Claude model ID

Changing only the model field can turn a working request into an HTTP 400. These are the documented breaks between the current tiers and from the legacy models, from Anthropic's what's-new pages for each model and its guide to choosing a Claude model:

• Forced tool use (tool_choice of type any or tool) returns 400 on Opus 5.5 and Sonnet 5.5 with tool_choice: type "tool" and "any" are not supported for this model. Use auto with strict tool use instead.
• Turning thinking off with thinking: {"type": "disabled"} returns 400 on Opus 5.5, where thinking is always on, and on Sonnet 5.5, which accepts {"type": "between_tools"} at low, medium, or high instead. Haiku 5.5 can turn thinking off at high or below.
• Manual budget_tokens returns 400 on all four current models; use effort instead.
• Moving from Haiku 4.5 to Haiku 5.5, non-default temperature, top_p, or top_k returns 400, and assistant prefill returns an error.
• The computer_20251124 computer use tool is rejected by Opus 5.5.
• Escalating mid-conversation keeps reasoning only in one direction: on the Claude API and Google Cloud, Opus 5.5 reads thinking blocks from Haiku 5.5 and Sonnet 5.5, but Haiku 5.5 doesn't read Opus 5.5's blocks and Sonnet 5.5 doesn't read Opus or Fable blocks. Unreadable blocks are dropped silently, and the request still succeeds.
• Switching with /model in Claude Code mid-session re-reads the conversation without cache hits, so the next turn costs more.

The chart shows the August 2026 lineup; the list above is current as of October 26, 2026.

Keep the model ID in configuration rather than scattered through code, so you can switch back without a release. Every Claude model ID is a pinned snapshot, including dateless IDs like claude-opus-5-5, so an ID never moves to a newer model on its own.

Test your own prompts before you change the default

No benchmark knows your workload, so the decision that sticks comes from running your prompts on two adjacent tiers. For app users, Anthropic's Academy tutorial suggests running the same prompt on Haiku and then Sonnet, such as "Summarize the attached report in five bullet points, then flag the one claim in it you'd most want a second source for," and comparing where the answers differ, not how long they are.

For the API, take a small set of prompts that cover routine cases, hard cases, and the failures that cost you most, and write the pass criteria before you look at any answer. Run the cheaper candidate and the next tier up with the same prompts, tools, effort policy, and maximum output, and record input and output tokens, time, retries, and whether each answer passed. Then compare cost per accepted answer rather than price per token:

cost per accepted answer =
  (API charges + retry charges + valued repair time) / accepted answers

Keep the cheaper tier if it passes. That is the same rule the October 26 test followed: when three tiers pass identical checks, the per-task bill decides, and there Haiku did the same passing work for about a tenth of Sonnet's cost.

Claude model choice: other questions

Which is better for coding, Sonnet or Opus?

For complex agentic coding, Opus 5.5. Anthropic's model guide recommends Opus 5.5 for multi-hour autonomous coding agents and large refactors, and Sonnet 5.5 for everyday code generation and agentic tool use. In Anthropic's own launch figures on Terminal-Bench 4.0, Opus 5.5 scored 66.4% at xhigh, Fable 5.1 55.8%, and Haiku 5.5 39.2%; those come from different launch posts and effort settings, so they aren't one matched test. For routine coding in the Claude apps, Sonnet 5.5 is Anthropic's suggested starting point.

Which Claude model is the most intelligent?

It depends on whose measure you use. Anthropic calls Fable 5.1 its most capable model open to all customers and recommends it for hours-long agent sessions and multistep research. On Artificial Analysis's Intelligence Index as of October 26, 2026, Opus 5.5 at max effort scores highest of the four at 58, ahead of Sonnet 5.5 at 56 and Fable 5.1 at 53. The index averages 10 evaluations and may not capture the hours-long agent work Anthropic positions Fable for.