Skip to content
senn-techsenn-tech
Back to blog
AI News
AI explained: models, market, news2026-11-15· By Franz Senn

Which Claude model? Anthropic's guidance checked against measured cost per task

A graphic circulating as Anthropic guidance seats the four current Claude models: Opus 5.5 as the daily driver for coding and knowledge work, Sonnet 5.5 as the saving option, Fable 5.2 for the open-ended work, Haiku 5.5 for the volume running behind the scenes. The lines are quick to read. Which currency they calculate in is not on the graphic.

Cost per task means the sum of every token one actually completed task consumes, billed at the chosen effort tier. The list price per million tokens says little about that, because token count, tier and tokenizer vary from job to job. Two posts on this blog evaluated it in September and October 2026: the same model type can cost double at a different setting, and the same model can stay expensive per task even while its token price falls.

Anthropic Claude trademark
Anthropic makes the four models compared here and maintains the pricing docs we checked list prices against. (Quelle: Wikimedia Commons, classed public domain (copyright); trademark of Anthropic)

What the guidance says

The wording as it was passed to us with the graphic (we have not seen the graphic itself, so the lines are unverified):

In general, use Opus 5.5 as your daily driver for coding and knowledge work. Switch to Sonnet 5.5 if you want faster replies and more runway with your usage. For your most ambitious, open-ended technical and research work, reach for Fable 5.2. If you're building on the API, Haiku 5.5 is our fastest and lightest model for high-volume work behind the scenes. These are rough guidelines, not rules—the best model is the one that fits the way you work.

The graphic is provider marketing material and is not embedded here; it reached us as a tip on 11 November 2026; we have not seen it ourselves and do not know where it comes from. We did not find this wording in the selection guide on claude.dev of 9 November 2026. That guide lists the same four models at the same list prices and recommends starting most work on Opus 5.5 at its default effort level, medium. The model documentation says the same in one sentence: "start with Claude Opus 5.5 for most workloads." We checked what can be checked: all four models exist, the list prices match Anthropic's pricing documentation on 15 November 2026, and the ranking fits what this blog has already written about Sonnet 5.5, Haiku 5.5 and Fable 5.2.

Runway is not a token bill

The recommendation speaks the language of subscription plans with usage quotas: switch to Sonnet 5.5 and you get more answers per quota. The plan pricing page does not mention an unlimited plan; the word "unlimited" does not appear there on 15 November 2026, and the sentence "Usage limits apply." appears several times. For the chat customer that is the right bill. A company running Claude through an API or its own gateway pays in tokens instead, and three levers the image never names take effect in that bill:

  • The effort tier. Anthropic promises Sonnet 5.5 up to 30 percent lower cost per task on tightly scoped work. Artificial Analysis measured 7.60 dollars per task on the highest tier on 28 September 2026, roughly 50 percent more than the predecessor Sonnet 5, on around 193,000 output tokens per task, at the time the highest figure that shop had measured (as reproduced in our Sonnet post). The documentation lists "high" as the default tier for Sonnet 5.5, and the claude.dev guide names "medium" for Opus 5.5.
  • The tokenizer. According to Anthropic's pricing documentation, which we read on 15 November 2026, Claude 4.7 and all later models use a newer tokenizer that produces roughly 30 percent more tokens for the same text; the exact increase depends on the content. A model on the newer line can therefore bill more tokens for the same text while nothing about its capability changes.
  • The cache. Per the pricing documentation, cache reads sit between 2.5 and 10 percent of the input price, depending on the model, and take effect only on unchanged, repeating prompts. How large the share is in your own gateway has to be measured there. The log of our gateway, quoted in the Fable post (176,380 requests over 30 days up to 9 November 2026), shows no hits for the LiteLLM response cache. The update dated 15 November in that post explains it: the vLLM prefix cache on the main lane reports 88.3 percent cached input tokens, a rate a cache at Anthropic could plausibly reach too (a calculation there, not a measurement; see the Fable 5.2 post). The zero is therefore no evidence against the cache.

The price ladder and the measured price per task

ModelList in/out, USD per million tokensIndex (AA)Cost per index task, max tier, 2026-11-15Earlier value in our posts
Haiku 5.50.10 / 0.50 (above 100k prompt: 0.50 / 2.50)36$0.41$0.21 (2026-10-07)
Sonnet 5.52.00 / 10.0055$10.02$7.60 (2026-09-28)
Opus 5.54.00 / 20.0059$10.04no figure
Fable 5.28.00 / 40.00 (fast mode 16 / 80)65$26.54$26.54 (2026-11-09)

List prices: Anthropic's pricing documentation, read on 15 November 2026. Cost per task and index: the four model pages of Artificial Analysis (AA), retrieved the same day; the pages label Sonnet, Opus and Fable "Max, Default Fallback" and Haiku "Max". The values have moved since launch, Sonnet from 7.60 to 10.02 and Haiku from 0.21 to 0.41. The page does not say why.

At max effort one index task therefore costs almost the same in total on Sonnet 5.5 and Opus 5.5, although the list asks double for Opus. AA gives 335 million output tokens for the index on Sonnet and 218 million on Opus, so Sonnet needs more tokens for the same tasks. Counting only output at list price gives 335 × 10 = 3,350 US dollars for Sonnet and 218 × 20 = 4,360 for Opus (derived); that the totals still come out level would have to be explained by input and cache costs or a different basis. The page gives no breakdown and we did not recompute one. By list price Haiku sits at a twentieth of Sonnet's price, per index task it is 0.41 against 10.02, about a twenty-fourth (derived). The graphic moves almost entirely through the flat, expensive top of the ladder.

Cost per index task at Max effort: Sonnet and Opus almost equalHaiku 5.50.41 · 0.41 USD, index 36Sonnet 5.510.02 · 10.02 USD, index 55Opus 5.510.04 · 10.04 USD, index 59Fable 5.226.54 · 26.54 USD, index 65030
AA measures Max effort; nobody there has evaluated the default levels. Values from Artificial Analysis, not re-measured by us. (Quelle: Artificial Analysis, model pages, retrieved 15 Nov 2026)

Three consequences for your own routing

The ranking is not wrong; as a routing rule it is incomplete. From the measured values three rules follow:

  1. Measure the price per task before the list decides. By list price Opus 5.5 costs double Sonnet 5.5, yet at max effort a task costs about 10 dollars on both, with Opus four index points ahead (output cost alone would be higher on Opus, see the table). The GDPval figures sit practically side by side, 1,846 versus 1,844 points on GDPval-AA v2.1, according to the vendor's page. Two caveats from our own Sonnet 5.5 post apply: the run used a pre-release build with a bug that could degrade answers, and the Opus figure appears on the vendor's page as 1,735 in the table and 1,846 in the running text, with no effort tier given. That two-point difference carries no decision. Starting on Opus 5.5 at the medium tier, as Anthropic recommends, is defensible on this data; whether Sonnet 5.5 comes out cheaper per task depends on tier and task. AA measures Max, and nobody there has evaluated the default tiers. Anyone who wants to know runs the same task on both models at the same tier. For speed, AA measured 138.7 tokens per second on Sonnet 5.5 on launch day; on 15 November the page shows 143 for Sonnet and 114 for Opus.
  2. Small models need counted consumption. By our Haiku post (as of 8 October 2026) Haiku 5.5 runs through around 162,000 output tokens per task at maximum effort, GPT-6 Luna around 50,000; at effort high both reach 38 points, Haiku on roughly 55,000 tokens, Luna on 50,000. At 0.10 dollars per million tokens Haiku sits at a twentieth of Sonnet's list price, and the consumption per task can claw back more of that than the name promises.
  3. Access questions come before price questions. Fable 5.2 costs 8 to 40 dollars, twenty percent under its predecessor Fable 5.1 (10 to 50), yet access demands 30 days of retention; without it the API answers HTTP 400 according to our Fable post, which we have not retried here. Whoever routes Fable to exceptions settles the retention question before any price table, with Covered Models laid out in the provider's policy.

Anthropic's own closing line, that the best model is the one that fits the way you work, is the actual rule: what fits is decided by counting, and the count lives in your own logs. How a self-hosted AI stack with LiteLLM and vLLM delivers the token counts per model and task, measurement points included, is written up in that post; more on putting running model news into context sits in the AI-explained topic hub.

Open points: We have not seen the graphic itself; its origin and the quoted wording are unconfirmed, and the claude.dev guide does not contain the wording. Output cost alone (3,350 versus 4,360 US dollars) does not fit equal totals; the breakdown is missing. We do not know why the AA costs moved since launch, and AA measures only the max tier. The GDPval figures come from the vendor and carry the caveats named. The 162,000 and 50,000 tokens come from our Haiku post and were not re-measured here. HTTP 400 without retention and the gateway count (without a cache figure) come from our Fable post, and the explanation of the zero (response cache versus prefix cache) comes from there and was not re-checked here. We ran no measurement of our own of the models against each other.

Read next: coding-agent harnesses up to 26 points apart · Claude Fable 5.2: price and retention period

Further Reading

Status per source: verified = fetched and read by us on 15 November 2026, reported = taken from our own earlier posts, vendor-claimed = manufacturer statement, derived = calculated by us.

Questions?
Which Claude model should a company gateway default to?+

The list does not settle it. On the list, Opus 5.5 costs double Sonnet 5.5 (4/20 versus 2/10 dollars per million tokens), yet in the Artificial Analysis index one task on the max tier costs about 10 dollars for both (10.04 and 10.02, retrieved 15 November 2026), at 59 versus 55 index points. Anthropic itself recommends starting on Opus 5.5 at the medium tier. Which model is cheaper for your task only a measurement at the same tier shows.

Is Haiku 5.5 the cheap choice for subagents?+

On the price sheet, yes: 0.10 dollars per million input tokens up to 100,000 prompt tokens, five times that above the line. The measurement makes it a question of consumption: in early October Artificial Analysis counted around 162,000 output tokens per task at maximum effort, while the rival GPT-6 Luna settled for about 50,000. On 15 November the page shows 0.41 dollars per index task for Haiku 5.5, against 0.21 in October. Haiku stays cheap only for whoever knows tier and consumption per task.

Does Anthropic's model guidance apply to API customers?+

It is worded for subscription plans with usage quotas, where more runway means more answers per quota. The API bills in tokens, and there the effort tier, the tokenizer and the cache-hit rate make the bill; the ranking on the image plays no role. Anyone who wants to check has to put a measurement point into their own gateway.