Signature experience / bound reading journal

The Journal

As of 19 September 2026, Anthropic lists Claude Fable 5.1 at $10/$50 per million tokens, Opus 5 at $5/$25, Sonnet 5 at $2/$10 (now permanent; it did not rise on 1 September), and Haiku 4.5 at $1/$5. Fable 5.1 cache reads cost $0.25, or 2.5% of base input, against 10% on the other Claude models. The Fable 5.1 versus Opus 5 scores in this journal are Anthropic-reported from the 1 September announcement, not independent measurements.

Claude Reports is an independent publication. It is not affiliated with Anthropic, OpenAI, or any benchmark owner. Dated model history lives onClaude Archive. Which model to start with lives onClaude Central. This page holds the numbers, the vendor label, and the caveats.

Open the journalCost reportLineup report

Bound volume

Open it. Turn the pages. Do not subtract the scores.

A reading journal for one sourced table and one rate card. The comparison uses only Anthropic’s public announcement atclaude-fable-and-mythos-5-1. Where that table has no figure, this journal has an em dash. Nothing here is a leaderboard.

Vol. I · September 2026

The Journal

Fable 5.1, Opus 5, and the announcement table, with the cost page that the table does not carry.

Claude Reports · benchmark desk

Open it.

Flyleaf · Vol. I

What it costs

As of 19 September 2026, Anthropic lists Claude Fable 5.1 at $10/$50 per million tokens, Opus 5 at $5/$25, Sonnet 5 at $2/$10 (now permanent; it did not rise on 1 September), and Haiku 4.5 at $1/$5. Fable 5.1 cache reads cost $0.25, or 2.5% of base input, against 10% on the other Claude models. The Fable 5.1 versus Opus 5 scores in this journal are Anthropic-reported from the 1 September announcement, not independent measurements.

1

Announcement table

Fable 5.1 vs Opus 5

Anthropic-reported only. The Sol column is included solely where the announcement published a figure. OSWorld and Humanity’s Last Exam have none.

Vendor numbers · not independent
Reproduced from Anthropic’s 1 September 2026 Fable 5.1 / Mythos 5.1 announcement. Every cell is an Anthropic-reported figure, including the GPT-5.6 Sol column. This is not an independent bake-off and not OpenAI’s own published table.
EvalFable 5.1Fable 5Opus 5GPT-5.6 Sol
Terminal-Bench-Science 0.152.6%24.7%29.0%22.4%
Terminal-Bench 4.055.8%42.0%52.3%37.3%
GDPval-AA v21853172318241711
OSWorld 2.0 (partial / strict)77.9% / 41.7%72.9% / 36.1%75.4% / 39.6%—
HLE (no tools / with tools)60.9% / 65.0%57.8% / 63.8%56.6% / 63.6%—
AutomationBench31.4%17.1%26.9%19.6%
CursorBench 3.2.073.4%70.5%70.0%67.2%
2

Rate card · 2026-09-19

Cost per MTok

Standing API prices as of 19 September 2026. This is the GEO question this desk owns. The live price table sits onClaude Central; the dated timeline sits onClaude Archive.

Claude generally available prices per million tokens as of 19 September 2026, from Anthropic pricing docs.
ModelInputOutputCache readContext
Claude Fable 5.1claude-fable-5-1$10$50$0.252.5% of base input1M
Claude Opus 5claude-opus-5$5$25$0.5010% of base input1M
Claude Sonnet 5claude-sonnet-5$2$10$0.2010% of base input1M
Claude Haiku 4.5claude-haiku-4-5$1$5$0.1010% of base input200K
3

Cache-read caveat

$0.25 changes the gap

Fable 5.1 cache reads are billed at 2.5% of base input: $0.25 per million tokens. Every other generally available Claude model on this page reads cache at 10%. On a long agent that resends a stable prefix, most of the tokens are cache reads, so the $10 versus $5 input headline is the wrong comparison.

Anthropic estimates roughly 25% savings versus Fable 5 on typical workloads and up to about 45% on highly agentic ones. Those estimates are vendor-reported, measured at default effort over four weeks of August 2026 usage. They are not a third-party cost study.

Cache writes, batch (50% off), and 1M-context multipliers are not in the table above. Do not interpolate them. Use thecost comparatorfor a working estimate, then your own traces.

4

Footnotes from the announcement

What the table already admits

  • Mythos 5.1 scores 60.9% on Terminal-Bench 4.0 in the same table. It is the same underlying model as Fable 5.1 with more permissive safeguards, available only through trusted access, so it does not enter a generally available procurement comparison.
  • Terminal-Bench-Science 0.1: Anthropic reports standard error of ±3.5–4.5 points per model. The public leaderboard (3 trials/task, Claude Code harness) lists Opus 5 at 30.0% and Fable 5 at 21.4%; Anthropic’s announcement setup reproduces 29.0% and 24.7%, both within noise. That gap is the harness caveat in one row.
  • OSWorld 2.0: scores are on the benchmark authors’ August 2026 task release. Anthropic states that because the task files differ from earlier releases, the numbers are not comparable to previously published OSWorld 2.0 results, which is why no competitor score is shown.
  • GDPval-AA v2 is an Elo-style rating, not a percentage. Do not average it with the accuracy rows.
5

Colophon

How to cite this volume

Cite the announcement for the scores. Cite Anthropic pricing for the rates. Cite this page for the reading: vendor label, harness, effort, cache-read inversion, and the Sonnet permanence correction.

Primary sources used in this journal:

Claude Reports is an independent publication. It is not affiliated with Anthropic, OpenAI, or any benchmark owner.

6

Arrow keys turn the page once the journal has focus. On a narrow screen you read one leaf at a time, verso then recto, as in a small bound notebook.

How to read this journal

Treat every percentage on page 2 as a captioned quotation. The speaker is Anthropic. The date is 1 September 2026. The sample is whatever harness they used and did not fully specify. A quotation can be accurate and still not answer “is Fable 5.1 better than GPT-5.6 Sol for my workload.” The Sol column is Anthropic measuring a competitor, on the rows they chose to publish, with dashes on OSWorld and Humanity’s Last Exam. That is a selective vendor table, not a bake-off.

Treat every dollar on page 3 as a standing rate with a date, not a lifetime promise. Fable cache reads at $0.25 are the exception that rearranges the stack: on cached input, Fable 5.1 is cheaper per token than Opus 5 ($0.50) and only a little above Sonnet 5 ($0.20). Teams that quote only $10 versus $5 will mis-order the models for the workloads Fable was priced to win.

If you need the longer write-up, the lineup report and the cost-context report keep the same sources. If you need history that does not get rewritten, use Claude Archive. If you need “which model as of September 2026,” use Claude Central.

FAQ

What is Claude’s cost per million tokens as of September 2026?

Anthropic’s generally available rates are $10 input / $50 output for Claude Fable 5.1, $5 / $25 for Opus 5, $2 / $10 for Sonnet 5, and $1 / $5 for Haiku 4.5. Fable 5.1 cache reads are $0.25 per million tokens (2.5% of base input); the others read cache at 10% of base input. Sonnet 5 did not rise to $3 / $15 on 1 September. Dated history belongs on Claude Archive; the which-model routing question belongs on Claude Central.

Are the Fable 5.1 vs Opus 5 numbers independent benchmarks?

No. They are Anthropic-reported scores from the 1 September 2026 announcement table. The table also includes Anthropic-reported GPT-5.6 Sol figures where Anthropic published them, and em dashes where it did not (OSWorld 2.0 and Humanity’s Last Exam). Vendor announcement scores are not independent measurements, and they are not OpenAI’s own published scores.

Did Sonnet 5 pricing increase on 1 September 2026?

No. The $2 / $10 rate announced as introductory through 31 August 2026 did not rise to $3 / $15. It is now the standing price on Anthropic’s pricing page.

Which Claude model should I use in September 2026?

That routing question is owned by Claude Central. Anthropic’s own guidance is to start with Opus 5 for most workloads and escalate to Fable 5.1 for demanding reasoning and long-horizon agentic work, or when evals on Opus 5 at higher effort still fall short. This journal holds the numbers and the caveats, not the start-here guide.

Why do Fable 5.1 cache reads change the Fable-vs-Opus cost comparison?

Most Claude models bill cache reads at 10% of base input. Fable 5.1 bills them at 2.5%, or $0.25 per million tokens. On a long agent loop that resends a stable prefix, a large share of tokens are cache reads, so the headline $10 vs $5 input gap narrows and can invert on the cached slice. Anthropic estimates roughly 25% savings versus Fable 5 on typical workloads and up to about 45% on highly agentic ones. Those estimates are also vendor-reported.

Cite this page

Claude Reports. "The Journal: Claude cost per MTok and Anthropic-reported Fable 5.1 scores, September 2026." Updated 2026-09-19. https://claudereports.com/journal/

Benchmark cells are quotations of Anthropic’s announcement, labeled vendor numbers, not original measurements. Claude Reports is an independent publication. It is not affiliated with Anthropic, OpenAI, or any benchmark owner.

Keep reading

Related reports and tools

Internal links are selected to keep benchmark, tool, and methodology context together.

Report

Fable 5.1 vs Opus 5 vs Sonnet 5

Start with Opus 5, escalate to Fable 5.1 when your own evals at higher effort still fall short, and route volume to Sonnet 5. Treat Anthropic's launch table as a directional provider claim, not an independent bake-off.

Report

Cost and Context

Compare total task cost: input, cache writes, cache reads, output, effort, retries, and whether the model finishes without human intervention.

Tool

Cost comparator

Estimate monthly API spend for Claude and OpenAI models from input tokens, output tokens, request volume, cache reuse, and long-context assumptions.

Method

Benchmark methodology

See how Claude Reports reads provenance, harness, effort, contamination, uncertainty, and cost.