In October 2026, only Kimi K3 is still its vendor's current model; GPT-5.6 Sol and Claude Fable 5 both have successors. Pick by workload: Claude Fable 5.1 for long agentic work, GPT-6.1 Sol for the cheapest output tokens, and Kimi K3 when you need open weights.
Quick verdict (prices per 1M tokens, checked October 21, 2026):
- Pick Claude Fable 5.1 (Fable 5's successor) for long-horizon agentic coding and research, at $10 input and $50 output.
- Pick GPT-6.1 Sol (Sol's value successor) when output cost decides, at $2 input and $10 output.
- Pick Kimi K3 when you want to download and run the weights yourself, or pay $15 output through Moonshot's API.
- Keep GPT-5.6 Sol only for existing cyber workloads; OpenAI now lists it under "Cyber models" at a promotional $4 and $20.
| Decision axis | GPT-5.6 Sol line | Claude Fable 5 line | Kimi K3 | Winner |
|---|---|---|---|---|
| Current model | GPT-6 Astra, GPT-6.1 Sol | Fable 5.1 | K3 (no successor found) | n/a |
| Lowest output price | $10 (GPT-6.1 Sol) | $50 (Fable 5.1) | $15 | GPT-6.1 Sol |
| Open weights | No | No | Yes, Kimi K3 License | Kimi K3 |
| Context window | 1.05M (5.6 Sol; GPT-6 Astra 1.05M) | 1M (Fable 5.1) | 1M | GPT-6 Astra / 5.6 Sol |
The map shows each July model, what replaced it by October 21, and whether its weights are downloadable.
How we compared the three models
We compared the models on vendor pages only, read on October 21, 2026. That meant Anthropic's models overview, OpenAI's model docs, OpenAI's pricing page and changelog, Moonshot's K3 blog, and the Kimi K3 model card. We ran no prompts ourselves. The axes are what a buyer decides on: which model is current, what output costs, and whether you can self-host.
Which model leads now? GPT-6 Astra and Fable 5.1, not the July three
Neither GPT-5.6 Sol nor Claude Fable 5 is its vendor's top recommendation anymore. OpenAI's model docs now say: "If you're not sure where to start, use GPT-6 Astra, our flagship model for complex reasoning and coding." They add GPT-6.1 Sol "to balance intelligence and cost."
OpenAI's changelog says GPT-6.1 Sol targets "complex coding and professional work at a lower cost than GPT-6 Astra." OpenAI's changelog lists it in early October 2026. GPT-5.6 Sol still sells, but as a cyber model with promotional pricing "available at least through November 21, 2026."
On September 1, 2026, Anthropic launched Claude Fable 5.1 as "the successor to Claude Fable 5 for long-running agentic coding, knowledge work, and research." It kept Fable 5's $10 and $50 prices and cut cache reads to $0.25. Fable 5 no longer appears in the current lineup table.
Two corrections to the July version of this post. Fable 5's launch page attributes the "first to break 90%" analytics line to a customer, Izzy Miller, not to an Anthropic benchmark. And the 1M context and 128K output limits are listed for Fable 5.1, not stated on the Fable 5 launch page.
Is Kimi K3 the cheap pick? Cheaper than Fable, not the cheapest
Kimi K3 costs $3 per 1M input tokens on a cache miss, $0.30 on a cache hit, and $15 per 1M output, per Moonshot. That is less than a third of Fable 5.1's output price. GPT-6.1 Sol is cheaper still on output.
The bigger change since July is the weights. Moonshot's K3 blog promised them "by July 27, 2026"; the Hugging Face card now reads "We release the full Kimi K3 model weights under the Kimi K3 License." That is a custom license, not the "Modified MIT" this post claimed before. Moonshot describes a 2.8T-parameter model using 16 of 896 experts, and the card lists 104B activated parameters.
Moonshot is candid about rank. It says K3's "overall performance still trails the most powerful proprietary models," naming Claude Fable 5 and GPT 5.6 Sol. Its own scores include 67.3 on DeepSWE and 90.4 on BrowseComp, both vendor-reported.
Pricing: what each option costs per 1M tokens
Output price is where these models differ most, so the chart ranks them by it.
| Model | Input / 1M | Output / 1M | Cached input / 1M | Status, October 21, 2026 |
|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $50 | $0.25 (cache read) | Current Anthropic model |
| GPT-6 Astra | $10 | $50 | $1 | OpenAI flagship |
| GPT-5.6 Sol | $4 | $20 | $0.40 | Cyber model, promo price |
| Kimi K3 | $3 (cache miss) | $15 | $0.30 (cache hit) | Current, open weights |
| GPT-6.1 Sol | $2 | $10 | $0.10 | OpenAI value tier |
GPT-6.1 Sol's price covers prompts up to 272K input tokens. GPT-5.6 Sol's figure is the standard short-context price. Fable 5.1's output costs 3.3 times K3's and 5 times GPT-6.1 Sol's.
Sources disagree on GPT-5.6 Sol. Composio's July comparison and Future AGI's August table list $5 input and $30 output. OpenAI's pricing page now shows the $4 and $20 promo.
One third-party test exists. Composio ran 12 tool-use cases: Fable 7 of 12, K3 7 of 12, Sol 6 of 12. It estimated about $7.76 per case for Fable, $1.39 for K3 and $2.69 for Sol. Its costs use July list prices, including Sol at $5/$30, and test the July models, not Fable 5.1 or GPT-6.1 Sol. It is vendor-run, small, and used different harnesses, so treat it as a hint.
Which should you pick? Match the model to the workload
Start with the job, then check price. The current models are the ones to compare, because two of the July three are now superseded.
- Long agentic coding, research or knowledge work: Claude Fable 5.1. Anthropic recommends it for "demanding reasoning and long-horizon agentic work." Fable 5's launch page says it routed some cybersecurity, biology and chemistry requests to Claude Opus 4.8; Anthropic's pages we read don't say whether 5.1 does the same.
- Coding at the lowest output cost: GPT-6.1 Sol at $10 output, half GPT-5.6 Sol's promo $20.
- Self-hosting, fine-tuning or data that can't leave your servers: Kimi K3, under the Kimi K3 License.
- Hardest reasoning on OpenAI's stack: GPT-6 Astra, the flagship, at Fable 5.1's price.
- Existing GPT-5.6 Sol cyber pipelines: keep them while the promo runs, and plan a move.
For the vendor products around these models, see our Claude, ChatGPT and Kimi pages, or the ChatGPT vs Claude comparison.
Pick none of the three if you need a default model
Most teams don't need a frontier tier for everyday work. Anthropic's own advice is to "start with Claude Opus 5.5 for most workloads," at $4 input and $20 output, against Fable 5.1's $10 and $50.
Sonnet 5.5 costs $2 and $10, and Haiku 5.5 starts at $0.10 and $0.50. If you want an open-weight model but not a 2.8T one, our Claude vs DeepSeek comparison covers another option.
How we checked
- What we did: Read Anthropic's models overview and release notes, OpenAI's model docs, pricing page and changelog, Moonshot's K3 blog and pricing page, and the Kimi K3 model card, October 21, 2026.
- Hands-on: We did not test these tools hands-on for this article.
- AI assistance: Researched and drafted with AI assistance; fact-checked and approved by the Toolbit team before publishing.
- Last verified: October 21, 2026
