Prototype · closed beta
unOpen.AI
In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.
Technology
Claude.
Built for longer work.
Claude Fable 5.1 is the Claude release we lead with for complex coding, document review and long-running agent tasks. Start with the work your current model struggles to finish, then compare quality and the cost of a verified result.
All Model Families →Introduction transcript. Qubi: Some jobs are long: a whole codebase, a contract set, a task of many steps. We hand it the whole job, and every finding is tied back to its source. A test we held back checks the patch, so a tidy story alone can't pass. A reviewer signs off, and we log the fixes, the time and the spend.
Current release
Claude Fable 5.1
The Claude release this page leads with is Claude Fable 5.1, released 1 September 2026. When we last reviewed it, Anthropic's other current models were Claude Mythos 5.1, Claude Opus 5, Claude Sonnet 5.
Reviewed against Anthropic's own documentation on 19 September 2026 (source). We check the public model catalogues every week; when a new release appears, we review it against the provider's announcement and update this page.
LATEST MODELS
The newest Claude models.
What Anthropic has shipped most recently. Each has its own page with specifications, independent ratings at every reasoning setting and listed prices.
7 Oct 2026
Claude Haiku 5.5
- Intelligence
- 36.2
- Context
- 1M
- In / out per M
- $0.100 / $0.500
28 Sep 2026
Claude Sonnet 5.5
- Intelligence
- 54.2
- Context
- 1M
- In / out per M
- $2.00 / $10.00
22 Sep 2026
Claude Opus 5.5
- Intelligence
- 58.4
- Context
- 1M
- In / out per M
- $4.00 / $20.00
1 Sep 2026 · our lead
Claude Fable 5.1
- Intelligence
- 49.0
- Context
- 1M
- In / out per M
- $10.00 / $50.00
EVERY MODEL
All 15 Claude models, compared.
Sort by any column. Select a model for its full page.
| Claude Haiku 5.5 | 7 Oct 2026 | 36.2 | 235 | 1M | $0.100 | $0.500 |
|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 28 Sep 2026 | 54.2 | 135 | 1M | $2.00 | $10.00 |
| Claude Opus 5.5 | 22 Sep 2026 | 58.4 | 98 | 1M | $4.00 | $20.00 |
| Claude Fable 5.1 | 1 Sep 2026 | 49.0 | 72 | 1M | $10.00 | $50.00 |
| Claude Opus 5 | 24 Jul 2026 | — | — | 1M | $5.00 | $25.00 |
| Claude Sonnet 5 | 30 Jun 2026 | — | — | 1M | $2.00 | $10.00 |
| Claude Fable 5 | 9 Jun 2026 | — | — | 1M | $10.00 | $50.00 |
| Claude Opus 4.8 | 28 May 2026 | — | — | 1M | $5.00 | $25.00 |
| Claude Opus 4.7 | 16 Apr 2026 | — | — | 1M | $5.00 | $25.00 |
| Claude Sonnet 4.6 | 17 Feb 2026 | — | — | 1M | $3.00 | $15.00 |
| Claude Opus 4.6 | 5 Feb 2026 | — | — | 1M | $5.00 | $25.00 |
| Claude Opus 4.5 | 24 Nov 2025 | — | — | 200k | $5.00 | $25.00 |
| Claude Haiku 4.5 | 15 Oct 2025 | — | — | 200k | $1.00 | $5.00 |
| Claude Sonnet 4.5 | 29 Sep 2025 | — | — | 1M | $3.00 | $15.00 |
| Claude Opus 4.1 | 5 Aug 2025 | — | — | 200k | $15.00 | $75.00 |
Intelligence and coding: Artificial Analysis indices at the model's best reasoning setting, and its measured output speed. Context and prices: the OpenRouter catalogue list price in USD per million tokens. Released: the provider's date where Artificial Analysis records one, otherwise the date the catalogue first listed it. Refreshed weekly; not measured by QQuantum.ai. Select a heading to sort.
September 2026 update
Evaluate the longer task.
Anthropic launched Fable 5.1 and Mythos 5.1 on 1 September. Fable is generally available; Mythos shares the underlying model with different safeguards and restricted trusted access. Enterprise Frontier Safeguards are planned for a phased rollout later this autumn, not assumed available today. Anthropic’s release and availability notes.
The chart below reproduces one provider-published comparison. It helps identify a candidate for evaluation, but does not establish which model will complete your workflow most cheaply. Keep the benchmark version and tested configuration attached to any score you use.
Published benchmark evidence
Published score (%) · Higher is better · Linear scale from 0 to 100.0
- Claude Fable 5.1
- 55.8
- Claude Opus 5
- 52.3
- Claude Fable 5
- 42.0
- GPT-5.6 Sol
- 37.3
Anthropic’s launch-table configurations, not matched-cost tests. Fable 5.1 uses production safeguards; the source describes fallback models for some safeguard interventions. These are historical tested versions, not a ranking of all current models.
Source: Anthropic September release comparison. Checked 19 September 2026. Provider-reported, not independently measured by QQuantum.ai.
A useful first pilot
An investigation with an answer you can check.
Proposed coding pilot: supply a reproducible bug, the relevant repository and an isolated test environment. Ask for a root-cause explanation and a small patch. Hold back a regression test so that an attractive explanation alone cannot pass the evaluation.
Record successful fixes, introduced regressions, reviewer time and total model spend. Compare several effort settings against your current model on the same cases. Reject solutions that change the requirement, disable a test or use data outside the agreed scope.
For long document work, use the same discipline: every material finding should point to a source passage, and conflicting evidence should remain visible. Our managed agent runtime guide covers the surrounding execution workflow; the September briefing puts this release alongside other customer opportunities.
INDEPENDENT RATINGS
How Claude compares.
Third-party scores and listed prices, refreshed every week. Useful for building a shortlist; your own evaluation decides.
Composite index across their reasoning, knowledge, maths and coding evaluations · Higher is better · Linear scale from 0 to 66.0
- Gemini 4 Carbon
- 59.4
- Claude Opus 5.5
- 58.4
- GPT-6 Astra
- 56.9
- Claude Sonnet 5.5
- 54.2
- Claude Fable 5.1
- 49.7
- Muse Wonder Full
- 49.7
- Kimi K3.1 Pro
- 46.0
- Claude Haiku 5.5
- 36.2
This family's highest-rated models, highlighted, against the best-rated model of each other family we cover. One composite score cannot show cost, latency, data residency or how a model handles your task.
Source: Artificial Analysis. Checked 30 October 2026. Measured independently by Artificial Analysis, not by QQuantum.ai or the providers.
USD per million output tokens · The newest releases in the family · Linear scale from 0 to $56.00
- Claude Haiku 5.5
- $0.50
- Claude Sonnet 5.5
- $10.00
- Claude Opus 5.5
- $20.00
- Claude Fable 5.1
- $50.00
- Claude Opus 5
- $25.00
Input is priced separately: Claude Haiku 5.5 $0.10, Claude Sonnet 5.5 $2.00, Claude Opus 5.5 $4.00, Claude Fable 5.1 $10.00, Claude Opus 5 $5.00 per million tokens. Catalogue list prices; the provider's own rates, caching and batch discounts can differ.
Source: OpenRouter model catalogue. Checked 30 October 2026. Listed by OpenRouter, not measured by QQuantum.ai.
Specification
- Provider
- Anthropic, United States
- September release
- Fable 5.1: generally available; Mythos 5.1: restricted trusted access
- Weights
- Closed — API access only; weights are not offered for self-hosting
- Context window
- Large; the exact figure varies by model tier — check the current model documentation
- Access paths
- Anthropic API, or AWS Bedrock, Google Cloud Vertex and other cloud catalogues inside your own account and region
- Cost levers
- Model tiering, prompt caching on a stable prefix, batch endpoints, and selective extended thinking
- Where we use it
- Long-document reasoning, agentic tool use, and instruction adherence under multiple simultaneous constraints
Provider-published specifications as of 19 September 2026, not figures we measured. Model families move quickly — check the provider's own documentation before committing to a number.
Where it fits well
01
Long-document reasoning
Large context windows handled coherently, which suits contract analysis, codebase work and multi-document synthesis where the alternative is aggressive chunking and reassembly.
02
Agentic tool use
Reliable multi-step tool sequences with sensible recovery when a call fails. This is the capability that most distinguishes production-viable agents from demos, and it is where model differences show up most sharply.
03
Instruction adherence under constraint
Holding to a schema, a tone and a policy simultaneously across a long interaction. Valuable wherever output feeds a downstream system that will break on a deviation.
04
Enterprise deployment paths
Available through AWS Bedrock and Google Cloud Vertex as well as the Anthropic API, which frequently resolves data-residency and procurement questions using infrastructure agreements you already hold.
The practical levers
-
Model tiering.
The family spans a wide capability and price range. Routing straightforward requests to a smaller model and reserving the largest for genuinely hard steps is the single largest cost lever available, and it usually costs nothing in quality when the routing is measured rather than guessed. -
Prompt caching.
Where a large stable prefix repeats — system instructions, few-shot examples, a fixed reference document — caching materially reduces both cost and latency on cache hits. It requires designing context so the stable part is a genuine prefix, which is worth planning for rather than retrofitting. -
Batch processing.
For work that does not need to be synchronous, batch endpoints cut cost substantially in exchange for latency. Document processing, backfills and offline analysis rarely need a synchronous response. -
Extended thinking.
Additional reasoning effort helps on genuinely hard problems and wastes money on easy ones. Enable it selectively, based on measurement rather than by default.
Where we would choose something else
Where no data may leave your infrastructure at all, a hosted API of any kind is unavailable and the answer is open weights on your own hardware — with a real capability trade-off that we would rather state than paper over.
Where a task is high-volume, narrow and well-defined — classification, extraction against a fixed schema — a much smaller model, sometimes fine-tuned, will frequently match quality at a fraction of the cost. Using a frontier model for that work is a common and expensive default.
And where an application is already deeply embedded in another provider's ecosystem, the integration cost of switching can outweigh a modest capability advantage. That is a legitimate engineering judgement, not a failure of nerve.
We are not an Anthropic reseller and receive nothing for recommending them. The selection is made per workload, against your constraints, and we benchmark rather than assume.
Questions
Questions about Claude.
Is our data used for training?
Under the provider's commercial API terms at the time of writing, inputs are not used for training. We confirm the specific terms and any reduced-retention option in the architecture document, and where you consume Claude through Bedrock or Vertex the data handling follows your existing agreement with that cloud provider — often the fastest route through procurement because the paperwork already exists. Confirm against the provider's current terms; they change.
How do you handle model deprecation?
We pin a specific model version in production and treat an upgrade as a change with a regression gate: run the evaluation suite against the new version, review the diff, then migrate deliberately. Systems where the model is a swappable component behind an interface survive deprecation cycles; systems whose prompts are entangled with one version do not.
Claude or GPT?
It depends on the workload, and anyone answering confidently without seeing yours is guessing. We benchmark both — and often an open model too — against your actual task and data, and choose on measured results including cost and latency, not on general reputation. The honest answer is frequently that either would work and the decision comes down to price or an existing infrastructure relationship.
Can we run Claude in our own VPC?
Not as self-hosted weights — it is not available that way. What you can do is consume it through Bedrock or Vertex inside your own cloud account and region, which addresses many data-residency requirements — confirm against your own obligations. Where the requirement is genuinely that no data leaves your hardware, open-weight models are the only option and we scope accordingly.
