Opus 5 vs Opus 4.8 vs Fable 5 vs GPT-5.6 Sol vs Kimi K3: Which One Do You Actually Run?

Five frontier models are now in play, and every comparison post this week has made the same mistake: ranking them.

There is no winner, because these five are not competing for one job. The useful output is not a leaderboard — it is a routing table.

Here is how I actually decide, and what each one costs.

The price sheet first

The first thing worth noticing is the least discussed: Opus 4.8 and Opus 5 cost exactly the same. Anthropic's own wording is that Opus 5 "provides greatly improved performance for the same cost as its predecessor."

So if you are still defaulting to Opus 4.8, you are not saving anything. You are paying current prices for last generation's results. On agentic terminal work the gap is 43.3 versus 18.7 — not incremental, a different tier. Migrate.

Where each one genuinely wins

Claude Opus 5 — the default for agents

Cheaper and stronger on orchestration, tool-calling, computer use and workflow automation. That is an unusual combination and it makes Opus 5 the sensible default for most agent systems.

The number that best captures it: on AutomationBench at medium effort, Opus 5 scored 24% at $0.89 per task, beating both Fable 5 and Opus 4.8 at under half their cost. The cheap setting on the cheap model beat the expensive setting on the expensive one.

Claude Fable 5 — deep single-repository code

Fable 5 keeps a narrow, real edge exactly where you would expect the most capable model to keep one:

• -SWE-bench Pro: 80 vs Opus 5's 79.2
• -FrontierCode 1.1: 53.5 vs 53.4
• -DeepSWE v1.1: 69.7 vs 68.8
• -Humanity's Last Exam, no tools: 56.5 vs 56.3

These margins are thin — fractions of a point in three of four cases. But they cluster in one place: hard, sustained work inside a single large codebase. If that is your day job, the flagship premium is defensible. For everything else it is not.

Fable 5 is not being retired. Anthropic lists it active with no end date before June 9, 2027. Ignore any post claiming otherwise.

GPT-5.6 Sol — abstract reasoning, and the long-context trap

Sol wins two rows outright: ARC-AGI-2 at 92.5 (against Opus 5's 90.4) and DeepSWE v1.1 at 72.7 (against 68.8). On pure abstract pattern reasoning it is the strongest of the five.

The trap is in the pricing. Any request carrying more than 272,000 input tokens is billed at 2× input and 1.5× output for the entire request. If you work with large codebases or long documents, Sol's real cost is meaningfully above its sticker price. That penalty does not exist on Opus 5, Fable 5 or K3.

Kimi K3 — the one you can own

K3 is the cheapest of the five at $3 / $15, and the only one whose weights you can download and run yourself. It leads the web-design arena and places in the top tier on general intelligence.

The honest caveats: the tooling around it is less mature than Anthropic's or OpenAI's, it reasons for a long time before acting, and serving capacity has been a real constraint — Moonshot paused new signups days after launch citing a compute crunch.

But for regulated data, air-gapped environments, or simply not wanting a vendor between you and your product, open weights are a feature no amount of money buys elsewhere.

Claude Opus 4.8 — nothing, now

Same price as Opus 5, comprehensively beaten by it. Its only remaining role is as a pinned snapshot if you have evaluation suites tuned to its exact behaviour.

The trap in every comparison chart

Here is the thing almost nobody publishing a benchmark table this week mentioned, and it is in Anthropic's own system card at section 8.5.

During the agentic engineering evaluation, safety classifiers refused a portion of API calls and those calls fell back to Opus 4.8. For Opus 5 that was about 5% of calls. For Fable 5 it was 42% of calls, across a quarter of all trials. GPT-5.6 Sol triggered none.

I am not raising that as a gotcha — Anthropic published it themselves. I am raising it because of what it means for you: a benchmark score is measured on the calls that completed. Your production system has to handle the ones that do not. Retries, fallbacks and latency are real costs that never appear in a score or on a price list, and they vary enormously between models.

That is the single best argument for building a routing layer rather than picking a favourite.

The routing table

Two rules that outlast this comparison

Match model intelligence to task intelligence. Most production work does not need a frontier model at max effort. The cheapest correct answer is the goal, not the highest score.

Add tools before you add thinking. Anthropic's own finding, from the system card: "agentic tool-use is generally a more cost-effective method of scaling test-time compute than adaptive thinking by itself." When an agent struggles, the instinct is to let it think longer. The cheaper fix is usually a better tool.

Every model in this table will be superseded. Those two rules will not be.

Sources: Claude pricing and model documentation, and the Claude Opus 5 system card (Table 8.1.A, §8.5, §8.12) for all Anthropic figures. GPT-5.6 Sol and Kimi K3 figures are from published pricing and public arena results. Where two sources disagreed, I used the system card, since it carries the numeric table.

A practical note that sits underneath every row in that table: the slowest part of using any of these models is usually not the model. It is you typing the instruction. I dictate mine now — on-device, nothing uploaded — with a tool I built for exactly this: IndianWhisper.