Claude Opus 5 vs GPT-5.6 Sol vs Fable 5 vs Kimi K3: Which AI Model Should You Pick in 2026?

July 2026 was a busy month.

On 1 July, Claude Fable 5 became available worldwide again after a suspension linked to US export restrictions. On 9 July, OpenAI launched GPT-5.6 Sol. Kimi K3 arrived on 16 July. Anthropic unveiled Claude Opus 5 on 24 July.

Four frontier models in a little over three weeks.

One clarification about Fable 5: it was introduced on 9 June, suspended on 12 June, and restored on 1 July. Model availability now depends as much on political decisions as on technical ones. That belongs in your architecture choices.

On public benchmarks, the gaps between these four models are small.

Inside a real application, they become visible. A model that is excellent at coding can be too expensive for thousands of support tickets. Another posts the best overall score and disappoints on your own document base.

So there is no best model in 2026. There is a model that fits a given task, at an acceptable cost, with a level of risk you accept. That is the angle we take here.

Everything below was verified on 27 July 2026. Prices, limits and availability change fast.

The short answer: which model should you choose?

For most companies, Claude Opus 5 is the best default. It matches or beats Fable 5 on many tests, for half the price.

GPT-5.6 Sol is our first pick for development agents and for workflows already built around Codex, ChatGPT or the OpenAI API.

Claude Fable 5 remains a premium model for genuinely exceptional tasks: certain analytical, scientific or very long software jobs. Its price and retention constraints make it hard to generalise.

Kimi K3 is the competitor to watch: a one million token context, native visual understanding and the lowest price of the four. Its value in production will depend on whether its weights and licence actually ship. We cover it in depth in our dedicated Kimi K3 article.

Comparison table: Opus 5 vs GPT-5.6 Sol vs Fable 5 vs Kimi K3

The prices below are API prices per million tokens. They exclude cache discounts, fast modes and the actual number of reasoning tokens consumed.

Two details these prices hide: Opus 5 Fast mode runs about 2.5 times faster and doubles the price, and Kimi K3 charges $0.30 for cached input.

Price per token does not tell you cost per task. A model that reasons for a long time can consume far more than a pricier model that answers in one pass.

Claude Opus 5: the best default for business

Anthropic sums up its promise this way: intelligence close to Fable 5 at Opus range pricing.

In practice, Opus 5 sometimes does better than Fable 5.

On FrontierBench v0.1, a benchmark of complex tasks run in a terminal, Opus 5 scores around 43.3% in max mode, ahead of Fable 5, GPT-5.6 Sol and well ahead of Opus 4.8. The best published result reaches 44.4% at xhigh effort, which shows maximum reasoning is not always the most efficient setting.

Opus 5 offers five effort levels: low, medium, high, xhigh and max.

That matters beyond the technical detail. It lets you use one model across several categories of task. A simple extraction runs fine at low effort, while a code audit or a strategic analysis justifies a high one.

The choice is no longer just « Opus or Sonnet ». You now tune the trade-off between capability, response time and token consumption inside a single model.

On the Artificial Analysis Intelligence Index, Opus 5 scores 61, against 60 for Fable 5, 59 for GPT-5.6 Sol and 57 for Kimi K3. It also posts the best results on GDPval-AA v2 and AA-Briefcase, which measure the ability to produce complete professional deliverables.

Where Opus 5 is most convincing

Opus 5 is well positioned for multi-tool business agents, complex software development, code audits and migrations, and high value document writing.

Its main advantage is balance: half the price of Fable 5 per token, the same one million token context, and equal or better results on several benchmarks.

The limits of Opus 5

Opus 5 does not win everywhere.

Fable 5 keeps better factual reliability on some tests. Artificial Analysis also notes that Opus 5 answers more readily, and that on its knowledge benchmark this comes with a higher hallucination rate than Fable 5.

You still need to connect the model to verifiable sources, require citations and review outputs on critical cases.

Fenxi verdict: Opus 5 becomes the default frontier model for most business projects.

GPT-5.6 Sol: the best pick for Codex, tools and sub-agents

GPT-5.6 is not a single model. OpenAI organised this generation into three lasting tiers: Sol, the flagship, Terra, the middle ground, and Luna, the fast and cheap option.

Sol is the one that concerns us here.

GPT-5.6 Sol introduces reasoning levels up to max, plus an ultra mode in some OpenAI products. In ChatGPT Work, Ultra is available on Pro and Enterprise plans. In Codex, it starts at the Plus plan.

The difference between Max and Ultra is not simply thinking for longer. Ultra runs several sub-agents in parallel, splits the task between them, then synthesises their results. The Responses API offers the same behaviour in beta.

That helps when analysing a repository module by module, comparing several sources in parallel, or having a second agent review the first one’s work.

GPT-5.6 Sol leads in its own coding environment

GPT-5.6 Sol scores 80 on the Artificial Analysis Coding Agent Index when used with Codex. It posts the best score on all three evaluations in that index, with a cost per task below Fable 5 in Claude Code.

Read that result carefully. An agent benchmark rarely measures the model alone: it also measures the system prompt, the tools and the execution environment. GPT-5.6 Sol benefits from Codex here, which guarantees nothing inside your own orchestrator.

The limits of GPT-5.6 Sol

Its output price, at $30 per million tokens, is higher than Opus 5.

Max and Ultra modes also increase processing time and volume consumed. Explicit cache charges writes at 1.25 times the normal input price, although cached reads keep a 90% discount.

Ultra should therefore be reserved for tasks whose value justifies several agents. Running five sub-agents to summarise an email is not architectural sophistication, it is waste.

Fenxi verdict: GPT-5.6 Sol is our first pick for teams already running Codex and for workflows that need advanced orchestration of tools and sub-agents.

Claude Fable 5: maximum power, harder to justify everywhere

Fable 5 forms a new class of model at Anthropic, sitting above the Opus range. At launch, Anthropic presented it as its most powerful generally available model, particularly strong in software development, professional analysis, vision and scientific research.

Fable 5 targets long tasks, the ones that need several hours of continuous work: large scale code migrations, autonomous agents running for hours or days, analysis of complex document bases, scientific research, and operations that require many successive checks.

Anthropic reports that a test at Stripe handled in a single day a migration that would have taken a team months of manual work. That is a launch figure, not a result that generalises to every project.

Why Fable 5 is no longer the obvious choice

Opus 5 directly challenges its positioning.

Fable 5 costs $10 per million input tokens and $50 output. Opus 5 costs exactly half that, and beats it on several benchmarks tied to agentic coding and professional work.

Fable keeps an edge on some tests, particularly software engineering, factual knowledge and especially long tasks. But its premium now has to be demonstrated case by case.

Fenxi verdict: Fable 5 should become an escalation model. Use it only when internal tests show a concrete gain over Opus 5.

Kimi K3: the Chinese competitor you can no longer ignore

Kimi K3 is probably the most strategic model in this comparison.

Built by Moonshot AI, it claims 2.8 trillion parameters in a Mixture of Experts architecture, native visual understanding and a one million token context. We break down its architecture and benchmarks in a dedicated article.

Its positioning covers three uses: long running software development, document and analytical work, and tasks that combine code and vision.

Kimi K3 can use screenshots as feedback to improve an interface, a game, a CAD environment or a frontend application. That is a real asset for visual coding workflows.

Its API price is the lowest of the four: $3 input, $15 output, $0.30 for cached input. That is 40% below Opus 5.

Is Kimi K3 really open source?

Moonshot presents Kimi K3 as an open model and announced a full weight release by 27 July 2026 at the latest.

When we checked on 27 July, the official documentation still described the release as upcoming. No complete Kimi K3 repository had appeared on Moonshot’s official Hugging Face profile, unlike the K2 generations already available.

The distinction matters. Until the files, the licence and the inference instructions are published, Kimi K3 is a proprietary model accessible through an API, with an announced weight release. Not a self-hosted option you can use today.

And even once published, 2.8 trillion parameters have nothing in common with installing a small model on a standard server. The sovereignty that open weights bring is paid for in hardware, operations and security.

Fenxi verdict: Kimi K3 deserves serious testing, particularly for frontend, multimodal and large context work. Moving it into production in Europe requires more contractual and operational checks.

Cross benchmarks: do the scores tell the whole story?

Here is a summary of results published in July 2026.

No model leads everywhere. Opus 5 is ahead on the general index and agentic professional work, GPT-5.6 Sol on coding with Codex, Fable 5 on certain software engineering tasks. Kimi K3 follows closely, at a much lower price.

But a benchmark does not reproduce your environment. Scores depend on the effort level, the number of attempts, the tools provided, the exact model version and how failures are counted.

Vendors also choose which results they publish.

That does not make benchmarks useless. They help you build a first shortlist of models to test. They do not replace an evaluation on your own data.

We recommend building a set of 30 to 100 cases representative of your business, then measuring success rate, critical errors, average cost per task and latency.

The best model is the one that completes your tasks, not the one with the highest score in a launch deck.

Which model for which use case?

Two points apply to all four models.

Sending a million tokens on every request is rarely good architecture. For a document base you query regularly, targeted search, RAG or a cache strategy costs less and gives more stable results.

On budget, routing remains the best strategy: a small model handles simple requests, Opus 5 or GPT-5.6 Sol takes the hard ones, Fable 5 is called only as a last resort, and a human validates sensitive actions.

What we recommend at Fenxi

For a new project, we would start from this architecture:

• main model: Claude Opus 5;
• specialised coding: GPT-5.6 Sol in Codex;
• experimental or low cost model: Kimi K3;
• exceptional escalation: Claude Fable 5;
• a small complementary model for high volume tasks;
• automated evaluation and human validation for critical actions.

This avoids depending on a single vendor. It also lets you switch models when prices, policies or performance change, which in 2026 can happen within weeks.

Conclusion

Model choice comes after scoping the need: which task to automate, which data is processed, what cost per operation is acceptable, what happens when the model is wrong, and whether you need to switch vendor quickly.

Answering those questions usually eliminates three of the four models before you open a single benchmark.

Unsure between Claude, GPT and Kimi for your project?

Tell us about your use case, the data involved and the expected volume. We can help you compare models on your own tasks, estimate the real cost per operation and build an AI architecture ready for production.