Claude Fable, Opus, Sonnet & Haiku: Pricing and Performance Compared [2026]

The first question most people run into when they start using Claude is "Should I use Opus, Sonnet, or Haiku?" As of September 2026, there is also a tier above all three: Fable 5.1.

Compare output prices on the API and the spread is 100x, from $50 for Fable 5.1 down to $0.50 for Haiku 5.5; even between Opus 5.5 and Haiku 5.5 it is 40x (Haiku 5.5's price applies to prompts up to 100,000 tokens; above that, output costs $2.50 and the gaps shrink to 20x and 8x). But the most expensive model is not always the right one. For some jobs Haiku is the sensible choice, while other tasks really do need Opus or higher.

This article compares pricing, performance, and speed as checked against Anthropic's official documentation on October 12, 2026, and works through cost estimates for common use cases to help you pick the right model. Model names and prices keep changing, so this article puts what to base the choice on ahead of which model to pick.

1. Where the Four Models Fit

Sources: Anthropic, "Models overview", "Choosing the right model", and each model's page (checked October 12, 2026)

Which model counts as the "starting point" changes every time a new model ships. What stays the same is the structure: at the top of its official model list, Anthropic names the model to start with when you're unsure, and the top-tier model to move up to when that isn't enough. As of October 12, 2026, those are Opus 5.5 and Fable 5.1, and the tables in this article follow that guidance. Opus 5, the previous starting point; Sonnet 5, which Sonnet 5.5 replaced on September 28; and Haiku 4.5, which Haiku 5.5 replaced on October 7, have moved to the Legacy section of the list (older, still available). They haven't gone away: Anthropic has committed not to retire Opus 5 before July 24, 2027, or Sonnet 5 before June 30, 2027. For Haiku 4.5, the commitment only runs to October 15, 2026, so see the availability notes in Section 5.

The names map to literary length. An opus is a large, complex work, standing for the deepest and longest thinking; a sonnet (a 14-line poem) is moderate depth; a haiku is short and quick. Fable sits one tier above Opus, and Anthropic describes it as its most capable generally available model.

2. API Pricing Comparison

Prices can change. Always check Anthropic's official pricing page for the latest figures (this article reflects prices as of October 12, 2026).

Standard Pricing (per Million Tokens)

Source: Anthropic, "Pricing" (checked October 12, 2026). A 1-hour cache write costs 2x the input price. For Haiku 5.5, the prompt length counts all input tokens in a request, including cache reads and writes, and a request over 100,000 tokens pays the higher prices even for the part that hits the cache. Restricting processing to the US only multiplies every line item by 1.1 (Claude 4.6 and later models; on Haiku 5.5 it also applies to the over-100K prices; Haiku 4.5 is not covered). Pricing through Google Cloud (Vertex AI) is covered in How to Use Claude on Vertex AI.

The gap between the most expensive output, Fable 5.1 ($50/MTok), and the cheapest, Haiku 5.5 ($0.50/MTok), is 100x; between Opus 5.5 and Haiku 5.5 it is 40x. Sonnet 5.5 costs half of Opus 5.5 on input, output, and cache hits, and 20x Haiku 5.5. But once a single prompt to Haiku 5.5 goes over 100,000 tokens, every item costs 5x ($0.50 input, $2.50 output), so for workloads that feed a long document on every request the gap with Sonnet 5.5 narrows to 4x. The Legacy Haiku 4.5 ($1/$5) costs 10x Haiku 5.5 for prompts up to 100,000 tokens and still 2x above that, so Haiku 5.5 is cheaper either way.

When Sonnet 5 was announced on June 30, 2026, its pricing was described as an introductory $2/$10 through August 31, rising to $3/$15 afterward. In an update to the announcement on August 10, Anthropic made $2/$10 the permanent price and stated there would be no increase. Sonnet 5.5, released on September 28, launched with the same prices as Sonnet 5, including caching and batch rates, but on October 7 its cache-hit price alone dropped from $0.20 to $0.10 (input, output, batch, and cache-write prices are unchanged). Opus pricing held at $5/$25 from Opus 4.5 through Opus 5, and Opus 5.5 is the first to drop it, to $4/$20 (in Anthropic's price list, the earlier Opus 4.1 was $15/$75). In the Opus 5.5 announcement, Anthropic also says that at default settings it is 40% cheaper than Opus 5 on typical tasks, but that is Anthropic's own measurement, combining the per-token price cut (20% on input and output) with fewer tokens used per task. Whether your own work gets the same savings is something you need to confirm by measuring usage. The Sonnet 5.5 announcement's "up to 30% less per task than Sonnet 5" is the same kind of figure: the per-token price is unchanged, and Anthropic measured the savings from using fewer tokens.

What Is a "Token"?

API usage is billed in "tokens." According to Anthropic, 1 million tokens is about 555,000 words (about 2.5 million Unicode characters) with the new tokenizer used by Claude 4.7 and later models such as Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1, and about 750,000 words for the Legacy Haiku 4.5 and other earlier models. The new tokenizer counts the same text as roughly 30% more tokens; the Sonnet 5 announcement puts it at 1.0 to 1.35x depending on the content. The ratio also varies by language, and code or non-English text can differ from these word counts, so the reliable approach is to measure your own text with Anthropic's token counting API.

3. What About Subscriptions?

If you use a monthly subscription instead of pay-as-you-go API billing, the cost structure changes significantly.

Sources: Claude pricing page, Claude Code, "Model configuration" (checked September 29, 2026). In Claude Code, Team and Enterprise also default to Opus 5.5; before v2.1.280, Pro and Team standard seats defaulted to Sonnet 5, and Max and Team premium seats to Opus 5

Default models have changed with each new release. According to the official Claude Code documentation, the Opus 5.5 update on September 22, 2026 (v2.1.280) made Opus 5.5 the default on every paid plan, including Pro. For chat, the Sonnet 5 announcement (June 30, 2026) said Free and Pro default to Sonnet 5, and the Opus 5 announcement (July 24) said Max defaults to Opus 5, but the Opus 5.5 and Sonnet 5.5 announcements do not mention any change to the chat defaults. Check the model picker in the app to see the current default. In Claude Code, Fable 5.1 is never the default on any plan; it is used only when you select it yourself.

Subscriptions are not unlimited: they have usage limits, but up to those limits you pay a flat rate. Anthropic does not publish how much each plan allows in tokens or dollars, so primary sources can't tell you "how many times cheaper than the API" a subscription is. If you hit the limit, you can keep going at standard API rates by turning on extra usage credits. In the Opus 5.5 announcement, Anthropic also says it raised the "5-hour usage limits" for Pro, Max, Team, and seat-based Enterprise, but it does not say by how much.

4. Performance Benchmarks

Each time Anthropic releases a new model, it builds a comparison table against whichever models it wants to compare at the time. As a result, all four models never appear in a single table. Here we lay out each announcement's table separately (all figures come from Anthropic's official announcement tables; don't compare numbers across tables, because the test conditions differ).

From the Opus 5.5 announcement: vs. Fable 5.1 and Opus 5

Source: comparison table in Anthropic, "Introducing Claude Opus 5.5" (September 22, 2026). Opus 5.5 was measured at maximum effort (max), except Terminal-Bench 4.0 at xhigh; OSWorld 2.0 is marked "partial" in the table's notes

Sonnet 5.5 announcement table: compared with Sonnet 5 and Opus 5.5

Source: comparison table in Anthropic, "Introducing Claude Sonnet 5.5" (September 28, 2026). Opus 5.5's Terminal-Bench 4.0 score is at xhigh. Sonnet 5.5 scores 46.2% on FrontierCode at max, lower than at xhigh; the table's note attributes this to running a code-review routine more often, which led to more out-of-scope changes. OSWorld 2.1 is marked "partial" for all three

From the Opus 5 and Sonnet 5 announcements: using the previous-generation Opus 4.8 as a yardstick

Opus 5 and Sonnet 5 were never compared in the same table. However, both announcements include the previous-generation Opus 4.8, so we use it as a common yardstick.

Sources: comparison tables in Anthropic, "Introducing Claude Opus 5" (July 24, 2026) and "Introducing Claude Sonnet 5" (June 30, 2026)

The ordering is clear. Within Anthropic's tables, Opus 5.5 beats Fable 5.1 and Opus 5 on all six rows, Opus 5 is well ahead of the previous-generation Opus 4.8 (roughly 2x on Frontier-Bench), and Sonnet 5 lands slightly short of Opus 4.8 (6.0 points behind on SWE-bench Pro, 2.3 points on Terminal-Bench 2.1). Anthropic itself says in the Sonnet 5 announcement that its performance is close to Opus 4.8 at a lower price. In the Sonnet 5.5 announcement's table, Sonnet 5.5 beats Sonnet 5 on all six rows and even beats Opus 5.5 on Terminal-Bench 4.0 (70.6% vs. 66.4%), but falls short of Opus 5.5 on the other five.

That said, don't read these numbers directly as "real-world differences." In the Opus 5.5 announcement, Anthropic itself writes that at this level, benchmark gaps have become an unreliable guide to real-world gaps, and that in its own internal use, the difference between Opus 5.5 and Fable 5.1 feels smaller than the scores suggest. The Sonnet 5.5 announcement likewise says that even where benchmark scores are close, Opus 5.5 remains clearly stronger at complex, open-ended work that requires sustained judgment. The Opus 5.5 figures were measured at maximum effort, not at the default effort (medium) you would normally use. Even on the same OSWorld 2.0, Opus 5 scores 74.0% in the Opus 5.5 announcement's table but 70.6% in the Opus 5 announcement's table, which shows the conditions differ from table to table. What you should base the choice on is not the ranking in a table but evaluations run on your own work. On price, Sonnet 5.5 is half of Opus 5.5, and Fable 5.1 is 2.5x Opus 5.5.

We don't use the knowledge-work benchmark GDPval-AA v2 for cross-table comparison, because the same Opus 4.8 has different scores in the two announcements (1,615 in the Sonnet 5 table, 1,593 in the Opus 5 table). For Opus 5.5's benchmark conditions and the rows where it loses to GPT-6 Astra, see our Opus 5.5 release breakdown; for Opus 5's detailed benchmarks, see the Opus 5 release breakdown.

Haiku doesn't appear in any of these tables. Anthropic's guide to choosing a model places Haiku 5.5 in the lowest-latency, lowest-price tier and describes it as built for high-volume, latency-sensitive work such as classification, routing, extraction, and subagent tasks. For simple tasks you can expect good enough accuracy at the lowest cost. Fable 5.1 trails in the Opus 5.5 table, but Anthropic still calls Fable 5.1 its most capable generally available model and points to it as the next step when Opus 5.5 at high effort still falls short in your evaluations. Claude Haiku 5.5, which the Sonnet 5.5 announcement said would arrive "in the coming weeks," was released on October 7, 2026. The tables in this section come from each model's announcement, so we have not added a Haiku 5.5 row. How to read Haiku 5.5's announced performance, plus its pricing and switching caveats, are covered in our Claude Haiku 5.5 guide.

5. Speed and Context Windows

Source: Anthropic, "Models overview" (checked October 12, 2026). Speed is a relative rating among the four models; actual speed varies with prompt and output length and with effort. The Legacy Opus 5, Sonnet 5, and Haiku 4.5 are not in the list's comparison table (knowledge runs through May 2026 for Opus 5 and January 2026 for Sonnet 5; Haiku 4.5 has a 200K-token context window, 64K-token max output, and knowledge through February 2025)

Haiku's biggest strength is speed. It responds the fastest of the four, which makes it a good fit for real-time chatbots and subagents. Fable 5.1, by contrast, is the tier that may spend a long time on a single response to a hard request. According to Anthropic's announcement, Opus 5.5 already produces output more than 30% faster than Opus 5 by default, and with fast mode (research preview ended October 15, 2026; no application needed; available on the Claude API and on Microsoft Foundry hosted on Azure. From October 22, 2026, also available on Google Cloud, operated by Anthropic; official "Fast mode") its output speed can go up to 2.5x. The trade-off is double the price ($8/$40). Sonnet 5.5, according to Anthropic's announcement, also produces output more than 30% faster than Sonnet 5, making it the fastest Sonnet so far.

The 1M-token context window and 128K-token max output are the same for all four: Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5. Haiku 5.5 alone, however, costs 5x per token once a prompt goes over 100,000 tokens, so for work that fills the 1M window its price gap with Sonnet 5.5 gets smaller. On the Batch API, Opus 5.5, Sonnet 5.5, Haiku 5.5, and the Legacy Opus 5 and Sonnet 5 can output up to 300K tokens when you add the beta header.

Another difference is how long each model stays available. Anthropic has committed not to retire Fable 5.1 before September 1, 2027, Opus 5.5 before September 22, 2027, Sonnet 5.5 before September 28, 2027, or Haiku 5.5 before October 7, 2027 (the Legacy Sonnet 5 is June 30, 2027, and Opus 5 is July 24, 2027). For the Legacy Haiku 4.5, the commitment only ran to October 15, 2026, and that date has now passed. As of October 16, 2026, the model deprecations page still lists it as Active (no retirement announced), but Anthropic can now announce its retirement at any time. Retirement comes at least 60 days after the announcement. If you run a system on Haiku 4.5, check that page regularly and plan a move to Haiku 5.5.

6. Cost Estimates by Use Case

Here is roughly what real tasks cost at standard API rates (no caching or batch). The per-token prices are the same as in the Section 2 table, and every prompt in these scenarios is under 100,000 tokens, so Haiku 5.5 is at its lower price (the Legacy Sonnet 5 has the same input and output prices as Sonnet 5.5, so at the same token counts it costs the same).

Scenario 1: Generate a roughly 2,000-word article

Input: about 1,000 tokens; output: about 4,000 tokens

Scenario 2: Load a code file and refactor it

Input: about 10,000 tokens (code + instructions); output: about 5,000 tokens

Scenario 3: A chatbot answering 1,000 inquiries a day

200 input tokens and 300 output tokens per inquiry × 1,000 inquiries

For high-volume work like a chatbot, the gap between Haiku 5.5 and Opus 5.5 grows to about $199 a month. Switching the same workload from Opus 5 to Opus 5.5 saves $51 a month for the same token count, and switching from the Legacy Haiku 4.5 to Haiku 5.5 takes $51 a month down to $5.10. Meanwhile, the gap between Sonnet 5.5 and Haiku 5.5 is about $97 a month. If Sonnet 5.5 improves answer quality there is still a case for it, but it is now well worth checking first whether Haiku 5.5 is good enough. A practical design is to use Haiku when its answers are good enough and route only complex inquiries to Sonnet or Opus.

Three caveats about these estimates. First, the tables above assume the same token counts for every model. Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5 share the same new tokenizer, so the ratios among those four hold, but Haiku 5.5 counts the same text as about 30% more tokens than the Legacy Haiku 4.5, so the saving from moving from Haiku 4.5 to Haiku 5.5 is somewhat smaller than the table's one-tenth (about one-eighth for the same text). Second, thinking (adaptive thinking): Opus 5.5 and Fable 5.1 always have thinking on and it can't be turned off, and Sonnet 5.5 and Haiku 5.5 also think before answering by default (Sonnet 5.5 has a between_tools setting that answers without thinking up front). Thinking tokens are billed as output. How much the model thinks depends on effort, and the default effort is medium for Opus 5.5 and Haiku 5.5 and high for Fable 5.1 and Sonnet 5.5. Third, Haiku 5.5's 100,000-token threshold: once the total input of a single request, including cache reads and writes, goes over 100,000 tokens, every item in that request is billed at 5x. Before going to production, measure actual usage with your own requests and effort settings.

7. Which Model Should You Choose?

FAQ