
Frontier models compared (November 2026)
Scope: five current flagship models, compared on the axes that only matter before you sign: prompt caching, Batch discounts, ZDR, ZOA, training use and data residency. Per-token rates match our pricing table; anything we could not verify is marked not confirmed rather than filled in.
The five models
| Model (official name) | Vendor | Position | Details |
|---|---|---|---|
| Claude Fable 5.2 | Anthropic | Released 2026-11-09; successor to Fable 5.1 at 20% lower rates | Fable 5.1 deep dive |
| GPT-6 Astra | OpenAI | Flagship announced 2026-09-03 | GPT-6 Astra deep dive |
| Gemini 4 Carbon | Announced 2026-10-28. The pricing page lists input $2.00 / output $10.00 (no free tier) | Pricing / Timeline | |
| Kimi K3.1 Pro | Moonshot AI | Live 2026-10-23; weights published 2026-11-04 | Kimi K3.1 Pro deep dive |
| Meta Muse Wonder | Meta | Launched 2026-10-12; API exposes muse-wonder-1.0 only | Timeline |
The larger Muse Wonder Full is available only through Muse Code subscriptions, not the public API, so it is outside this comparison.
Three things to take away
- Cache read discounts span 4x across the five models. The deepest is Claude Fable 5.2 at 0.025x (97.5% off); Muse Wonder sits in the middle at 0.05x (95% off); the shallowest is 0.1x (90% off) on GPT-6 Astra and Kimi K3.1 Pro. For a workload that resends the same prefix, the multiplier can matter more than the headline input rate.
- Discount rate and absolute price are different questions. The most expensive cached read here is GPT-6 Astra at $1.00; Fable 5.2, Kimi K3.1 Pro and Muse Wonder all sit at $0.20.
- Choosing ZDR costs you features. Meta states that with ZDR enabled, server-side conversation state, background requests, file uploads, web search and Contributor mode are unavailable. Retention settings and feature settings are the same switch.
1. Rates (USD per 1M tokens, standard tier)
| Model | Input | Output | Cache read | Write 5 min | Write 1 hour | Notes |
|---|---|---|---|---|---|---|
| Claude Fable 5.2 | $8.00 | $40.00 | $0.20 0.025x | $10.00 1.25x | $16.00 2x | Fast mode $16 / $80 |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 0.1x | Not confirmed | Not confirmed | Long tier $20 / $2 / $75. Fast mode 2x |
| Gemini 4 Carbon | $2.00 $4.00 from 1 Dec | $10.00 $20.00 from 1 Dec | $0.10 $0.20 from 1 Dec | Not listed | Not listed | Implicit caching only (no explicit caching) |
| Kimi K3.1 Pro | $2.00 | $10.00 | $0.20 0.1x | $2.00 | $4.00 | Highspeed tier is 2x |
| Meta Muse Wonder | $4.00 | $20.00 | $0.20 95% off | Not confirmed | Not confirmed | No long-context premium |
Fable 5.2 cache rates were read off the Claude pricing page with the column headers intact (fetched 2026-11-14): 5m writes $10, 1h writes $16, hits and refreshes $0.20. Our earlier "not confirmed" came from a summary table showing "Prompt caching $40", which looked like a 5x marker; the breakdown is now resolved.
2. Cache mechanics (minimum length, retention, mode)
| Model | Mode | Minimum cacheable | Retention (TTL) | What it means |
|---|---|---|---|---|
| Claude Fable 5.2 | Explicit (cache_control) | Not confirmed | Choose 5 minutes or 1 hour at write time | The only one of the five where you pick the TTL |
| GPT-6 Astra | Automatic + explicit | 1,024 (stated for the GPT-6.1 line; Astra-specific value not confirmed) | 30 min default; 12 h option in the Responses API (GPT-6.1 line) | Works on short prefixes, but the read rate is the highest here |
| Gemini 4 Carbon | Implicit only | Not listed | Not listed | No explicit caching means you cannot set the boundary yourself |
| Kimi K3.1 Pro | Automatic by default | Not confirmed | 5 min default; each hit refreshes the TTL | No configuration needed; weak for long-lived reuse |
| Meta Muse Wonder | Automatic | Not confirmed | Not confirmed | Under ZDR, server-side state is unavailable entirely |
For explicit vs implicit caching, see the glossary entry; per-model rates are in Prompt caching rates compared.
3. Batch discounts and combining with cache
| Model | Batch discount | Batch rate (calculated) | Combine with cache? |
|---|---|---|---|
| Claude Fable 5.2 | 50% off | $4.00 / $20.00 | Yes — "Batch API and prompt caching discounts can be combined." |
| GPT-6 Astra | 50% off | $5.00 / $25.00 | Not confirmed |
| Gemini 4 Carbon | Not supported | — | — ("Batch API isn't supported") |
| Kimi K3.1 Pro | 60% of standard | $1.20 / $6.00 | Not confirmed |
| Meta Muse Wonder | Not confirmed | — | — |
Turnaround, retention windows and limits are in Batch API discounts compared. Batch rates here are calculated (standard rate x discount), not figures each vendor published separately.
4. Data controls (ZDR, ZOA, training, residency)
| Model | ZDR | ZOA (operator access) | Training (API default) | Data location |
|---|---|---|---|---|
| Claude Fable 5.2 | On request Fable 5.1 and earlier are Covered Models requiring 30-day retention and are excluded from ZDR; whether 5.2 is on that list is not confirmed | Controlled access path Customer-managed keys not confirmed | Commercial terms: no training on Customer Content | No region choice on the first-party API (not even EU) |
| GPT-6 Astra | Approval required Zero Data Retention or Modified Abuse Monitoring customers only | Safety review runtime designed to disable human access; EKM | Not used for training by default on the API | Storage selectable; processing is not (Japan: storage yes, processing no) |
| Gemini 4 Carbon | Available on Google Cloud customer-side configuration required | Access Approval / Access Transparency / KAJ | Paid tier: not used. Free tier (Unpaid Services) is used for improvement with human review | Google Cloud supports region selection; AI Studio location not confirmed |
| Kimi K3.1 Pro | Not confirmed | Not confirmed | Not confirmed | Not confirmed |
| Meta Muse Wonder | Approval required qualified Meta Model API accounts; ZDR disables stateful requests, background requests, file uploads and web search | Not confirmed | Not confirmed Contributor mode is unavailable under ZDR per the official help centre | Not confirmed |
ZDR, ZOA, training and residency are separate documents. Background: ZDR / ZOA / training use / data residency. Satisfying one does not satisfy another.
5. Context window and output ceiling
| Model | Context window | Max output | Long-context billing |
|---|---|---|---|
| Claude Fable 5.2 | Not confirmed Fable 5.1 is 1M | Not confirmed 5.1 is 128K | No tiered pricing (per 5.1) |
| GPT-6 Astra | 1.05M | 128K | Long tier $20 / $2 / $75 |
| Gemini 4 Carbon | 1M tokens | 64K / 1M tokens | No long-context tiering listed on the pricing page (source: ai.google.dev/gemini-api/docs/models, retrieved 2026-11-15) |
| Kimi K3.1 Pro | 1,048,576 | Not confirmed | None listed |
| Meta Muse Wonder | 1M | Not confirmed | No long-context premium |
Cells we could not confirm
| Item | Model | Why it is missing |
|---|---|---|
| Minimum cacheable length | Fable 5.2 / Kimi K3.1 Pro / Muse Wonder | No minimum token count in the pricing pages we fetched, and none in Anthropic's caching docs either |
| Cache write rate | GPT-6 Astra / Muse Wonder | The official cache rate table has no row for these models |
| ZDR, ZOA, training, residency | Kimi K3.1 Pro | No verbatim quote obtainable from the official data-handling pages; third-party summaries exist but are not primary sources |
| ZOA and residency | Meta Muse Wonder | The official help centre covers ZDR only |
| Context window and max output | Claude Fable 5.2 / GPT-6 Astra | Not on the pricing pages; model docs not yet retrieved verbatim |
| Batch discount | Meta Muse Wonder | No Batch-equivalent statement found on Meta's pricing page |
| Availability stage (GA / preview) | Gemini 4 Carbon | Generally available (GA) on 2026-11-10. The pricing page itself carries no stage label, so the stage is recorded from the operator's confirmation (okamo); the rates come only from the paid-tier figures on that page |
"Not confirmed" does not mean "not available". We fill cells only from primary sources, and update them as we verify each one.
Sources (all fetched 2026-11-14)
- Anthropic — Pricing (Fable 5.2 $8 in / $40 out, 5m writes $10, 1h writes $16, hits $0.20, Batch 50% off)
- OpenAI — pricing page (GPT-6 Astra $10 / $50, cached input $1.00) and model page (1.05M context, 128K output)
- Google — Gemini Developer API pricing (Carbon $2.00/$10.00 rising to $4.00/$20.00 on 1 Dec; cache $0.10 to $0.20; implicit only) and the Gemini 4 Hydrogen announcement (Carbon: "Batch API isn't supported")
- Moonshot AI — platform.kimi.ai pricing page (K3.1 Pro $2.00 / $10.00, cache read $0.20, writes $2.00 for 5m and $4.00 for 1h, Batch at 60% of standard)
- Meta — Model API: Zero Data Retention ("ZDR is available for qualified Meta Model API accounts" plus the list of unavailable features) and the pricing page ($4.00 / $20.00, cached input $0.20, 1M context)
Related pages
- Prompt caching rates compared — write and read rates across five vendors
- Batch API discounts compared — discounts, turnaround and retention
- LLM API pricing — standard rates with per-model verification dates
- ZDR / ZOA / Data residency — the definitions behind the table above
- Token cost calculator — estimate from input/output token counts