Gemini 4 Carbon vs Claude Fable 5.1: The Premium Model Anthropic Already Retired

G​emini 4 Carbon went on general sale on the paid G​emini API on 10 November 2026, and the model it gets stacked against most often — Claude Fable 5.1 — is no longer the flagship of its own family. A​nthropic's model documentation now files Claude Fable 5.1 as "Active (legacy)", sends readers to Claude Fable 5.2 at $8 and $40 per million tokens, and commits to supporting 5.1 no sooner than 1 September 2027. That makes this an unusual head-to-head: a model one day into general availability, against a former flagship still sold at its original $10 and $50. The interesting question is not which is better. It is whether a legacy-priced $10/$50 model can justify a premium over a newer one that costs 40% as much per token on the independent board and scores eleven points higher.

The two envelopes, with the status field included

Spec sheets for these models usually leave out the field that matters most here, so it goes first.

• Status — G​emini 4 Carbon: generally available on the paid G​emini API, AI Studio and Android Studio since 10 November 2026, after a gated announcement on 28 October to cyber defenders and Fairwind Program partners. Claude Fable 5.1: released 1 September 2026, now marked Active (legacy), with Claude Fable 5.2 as the recommended replacement.

• Model ID — gemini-4-carbon versus claude-fable-5-1, the latter also carried as anthropic.claude-fable-5-1 on Bedrock and under the same string on G​oogle Cloud, Microsoft Foundry and C​laude Platform on AWS.

• List price — Carbon $2 in and $10 out per million tokens through 30 November 2026, then $4 and $20 from 1 December. Fable 5.1 $10 in and $50 out, with no announced change.

• Context window — 1,000,000 tokens each, and neither vendor applies a long-context premium. A 900,000-token request bills at the same per-token rate as a 9,000-token one on both.

• Maximum output — Fable 5.1 caps synchronous responses at 128,000 tokens. G​oogle's model documentation lists Carbon's output ceiling as 65K with long-decode generation up to 1M tokens, which is a less conventional shape and worth reading on the page rather than assuming.

• Input modality — Carbon accepts text, image, audio, video and PDF. Fable 5.1 accepts text and images. Neither produces anything but text.

• Caching — Carbon offers implicit caching only, at $0.10 per million cached input tokens until 30 November and $0.20 after; explicit context caching is not supported. Fable 5.1 bills cache reads at $0.25 per million — 2.5% of its input price — with cache writes at $12.50 for a five-minute window and $20 for an hour.

• Free tier — Carbon has none. Every token is billed, including thinking tokens. Fable 5.1 has no free API tier either; both are paid-only models.

One asymmetry in that list is easy to skim past. Carbon's discounted cache price is cheaper than Fable 5.1's, but Carbon cannot pin a cache explicitly, so the discount only arrives when G​oogle's automatic prefix matching happens to fire. Fable 5.1 lets you mark the breakpoint and guarantee the hit — which is what makes its 2.5% rate a budgetable number rather than a hope.

The 1 December cliff, and why it is the only date that matters

Carbon is on the same introductory rate card G​oogle used for Gemini 4 Argon, and the pattern is now familiar enough to plan against: $2 and $10 per million until 30 November 2026, then $4 and $20 for the same model with nothing else changed. Cached input moves from $0.10 to $0.20 on the same day, and batch and flex double from $1/$5 to $2/$10.

That schedule turns a straightforward price comparison into a calendar problem. Today, Carbon costs a fifth of Fable 5.1 on input and a fifth on output. From 1 December it costs two-fifths on both. Neither number is wrong; they are three weeks apart, and a workload priced on the first will look twice as good as it is.

The comparison that survives the calendar is the standard-rate one: $4 and $20 against $10 and $50. On that basis Carbon is 60% cheaper in both directions, with a cache read at $0.20 against $0.25 that is close enough to call even. If you are modelling a migration, use those numbers and treat November as a discount window rather than a price.

What the independent board says, at the settings it was run at

Artificial Analysis has measured both models, which is rarer than it should be in this price band, and the gap is wider than the price gap. Every figure in this section is that evaluator's, read on 11 November 2026, not either vendor's:

• Intelligence Index v5.0.1 — G​emini 4 Carbon at high effort scores 60, ranked fourth of 74 models. Claude Fable 5.1 scores 49, ranked 24th.

• Cost per index task — $7.76 for Carbon against $12.99 for Fable 5.1. The model that costs 60% less per token also costs 40% less per completed task.

• Output speed — 79 tokens per second for Carbon, 88 for Fable 5.1. Fable is the faster of the two, narrowly.

• Time to first answer token — 67 seconds for Carbon, 166 seconds for Fable 5.1. Neither is a chat model in any useful sense.

• Verbosity on the index — Carbon consumed 205 million output tokens, 70th of 73 on that measure; Fable 5.1 consumed 163 million. Both think for a long time before answering.

• Context — 1M tokens for both, which matches each vendor's own documentation.

The effort setting is where a careless comparison goes wrong. Carbon's dial moves its score: high is 60, medium 58, low 54. Fable 5.1 has five levels — low, medium, high, xhigh and max — with high as the API default, and its published benchmarks are curves across those levels rather than points. Artificial Analysis measured Fable 5.1 at max. So the 60-versus-49 line is high-effort Carbon against max-effort Fable, which happens to be the most favourable pairing for the older model, and it still loses by eleven points.

Where Fable 5.1 still earns its price

Three things, and none of them is raw capability.

The first is deployment surface. Fable 5.1 is on the C​laude API, Bedrock, G​oogle Cloud in both Google-managed and Anthropic-managed forms, Microsoft Foundry and C​laude Platform on AWS, all under one model ID family. A regulated buyer who needs the model inside an existing cloud contract is choosing between a vendor with five routes and a vendor with one paid API. Carbon's 10 November post lists the G​emini API, AI Studio, Android Studio, G​emini for business, Antigravity, CodeMender and the G​emini app — a broad list, but all of it G​oogle's own estate except where enterprises front it themselves.

The second is refusal behaviour, which is a feature if you have built for it. A​nthropic's safeguards can decline a request outright and return a normal HTTP 200 with a refusal stop reason rather than an error, and the documentation ships explicit fallback guidance for that path. Teams that already handle it get a predictable contract. Teams that do not discover it in production.

The third is the effort dial itself. Five levels with a documented default of high, and per-message effort changes that do not invalidate the prompt cache, give a cost engineer more places to stand than Carbon's three. Whether that is worth a 2.5x token price is a question about your workload, not about the models.

What it does not buy is a reason to stay. A​nthropic's own comparison table now leads with Claude Fable 5.2 at $8 and $40 — cheaper than 5.1 on both meters, same 1M context, same 128K output, same June 2026 knowledge cutoff — and tells 5.1 users to migrate. When a vendor discounts its successor below the model you are comparing, the premium has already been withdrawn; the rate card just has not caught up.

The decision, stated as arithmetic

Take a workload that costs $1,000 a month on Fable 5.1 at $10 in and $50 out, split evenly. Moving it to Carbon at the standard $4/$20 costs $400 for the same tokens. Moving it at the November rate costs $200. Moving it on 1 December costs $400, and any pilot you run this month will report $200 and mislead you.

The capability side runs the same direction. Carbon scores 60 to Fable's 49 on the board both are measured on, costs less per task as well as per token, takes audio, video and PDF input rather than text and images, and is one day into general availability rather than fourteen months into a legacy window. The only axis where Fable 5.1 leads is measured speed — 88 tokens per second against 79 — and 9 tokens per second does not close an eleven-point gap.

The honest counterargument is not about the models at all. It is that Fable 5.1 has a year of production behaviour behind it, a published system card, and a retirement commitment you can put in a risk register, while Carbon has been callable by the general public for a single day and G​oogle has published no rate limits. If your workload cannot absorb an unknown throughput ceiling, that is a real reason to wait a month — and waiting also means you will be pricing on the December rate rather than discovering it.

Running both without standardising on either

The practical obstacle to any of this is that the two models live behind two different contracts, two SDKs and two billing relationships, which is exactly the situation a routing layer exists to remove. anthropic/claude-fable-5.1 is on OrcaRouter at A​nthropic's own $10 and $50 with the $0.25 cache read, and anthropic/claude-fable-5.2 is there at $8 and $40, which makes the successor comparison a config change rather than a migration. G​emini 4 Carbon is not in the catalogue: it is reached through G​oogle's paid API, and nothing here should be read as claiming we serve it. What the routing layer does buy is a way to run the comparison — one key, an OpenAI-compatible surface, and a fallback chain so a pilot that starts on Carbon can fail over to Fable 5.2, or to a cheaper model entirely, without a second integration.

That matters more than usual in a week where both sides of the comparison moved. Carbon's access changed twice in fourteen days: gated announcement, then paid API, with a repricing already scheduled. Fable 5.1's status changed from flagship to legacy while its price stayed put. Building a single-vendor dependency into a month like this is a bet on both vendors holding still, and neither has.

What would change this answer

Two things, and both are watchable rather than speculative. If G​oogle publishes Carbon rate limits and they turn out to be tight, the throughput argument flips toward the model with a known ceiling. And if A​nthropic retires Fable 5.1 early — the commitment is "not sooner than" September 2027, not a promise to keep it that long — the comparison stops being a choice and becomes a migration with a deadline.

Until one of those happens, the ranking is not close. Carbon is the newer model, the cheaper model per token, the cheaper model per task, the higher-scoring model on the only board that has measured both, and the one with the wider input surface. Fable 5.1 is a well-documented $10/$50 model whose own vendor has already published something better for less. The premium is still on the rate card; it is no longer in the argument.