Since 1 September, almost every major lab has shipped a new model: Anthropic's Claude 5.5 family, OpenAI's GPT-6 and GPT-6.1, Google's Gemini 4 Argon, Meta's first frontier model and a wave of strong open-weight releases. The leaderboard changed more in six weeks than in the previous six months. This guide puts them side by side on independent benchmarks and on what they actually cost to run, then turns that into simple recommendations.
The leaderboard at a glance
The Artificial Analysis Intelligence Index combines ten demanding benchmarks covering reasoning, knowledge, maths and coding into one score. It is run independently on every model, which makes it a fairer comparison than the charts in each lab's launch post. Scores below are each model's best reasoning setting.
Source: Artificial Analysis leaderboard, mid-October 2026. Best available reasoning setting for each model. Muse Wonder Full is only available inside Meta's Muse Code subscription.
Two things stand out. First, the top eight models sit within six points of each other, so for most everyday work any of them is "good enough" and price, speed and availability matter more than the ranking. Second, the gap between the best closed model and the best open-weight model is now about 12 points, which in practice means open models are roughly where the frontier was in the spring.
Quality versus cost: the chart that matters
The fairest way to compare price is cost per task: what it actually costs to run the benchmark, including every reasoning and output token. Hover over or tap a dot for details, and use the legend to hide groups. The best models are towards the top left.
Cost per task is Artificial Analysis's cost to run its index on each model, mid-October 2026. Muse Wonder Full is not shown because it has no public API price.
Read it like this: GPT-6.1 Sol and Gemini 4 Argon (at its introductory price) deliver close to top-tier scores at a fraction of the cost of the Claude models at maximum effort. MiMo-V2.6-Pro and Claude Haiku 5.5 are the stand-out cheap options. And Grok 4.7 is an example of why per-token price misleads: at $2 input and $6 output per million tokens it looks cheap, but it produces so many tokens that each task costs more than GPT-6.1 Sol.
Price per million tokens
For budgeting API use, these are the list prices for input and output tokens. Cached input, which you get when you resend the same long prompt or document, is far cheaper on every platform.
| Model | Input | Output | Notes |
|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | 1M context, no long-context surcharge |
| Claude Sonnet 5.5 | $2 | $10 | 1M context, no long-context surcharge |
| Claude Haiku 5.5 | $0.10 | $0.50 | Prompts up to 100K tokens; $0.50 / $2.50 above that |
| GPT-6 Astra | $10 | $50 | OpenAI's flagship |
| GPT-6.1 Sol | $2 | $10 | Cached input $0.10 |
| GPT-6 Luna | $0.10 | $0.50 | Cheapest model tracked |
| Gemini 4 Argon | $2 | $10 | Introductory until 30 November, then $4 / $20 |
| Meta Muse Wonder | $4 | $20 | Via the Meta Model API |
| Grok 4.7 | $2 | $6 | Uses many output tokens |
| MiMo-V2.6-Pro | $0.435 | $0.87 | Open weights, MIT licence |
The contenders
Claude Opus 5.5
#1 overall- Index
- 58
- Price
- $4 / $20
- Released
- 22 September
- Highest independent score, strongest on long agentic tasks
- 1M-token context at standard price
- Most expensive per task at max effort
Pick it when quality matters more than cost.
Claude Sonnet 5.5
Best all-rounder- Index
- 56
- Price
- $2 / $10
- Released
- 28 September
- Within two points of Opus at half the token price
- Second-highest independent score, behind only Opus
- Max effort uses a lot of tokens
The default choice for most teams.
GPT-6.1 Sol
Best value- Index
- 52
- Price
- $2 / $10
- Released
- 29 September
- Near-Astra results at about a fifth of the cost
- 1.05M context, 128K output
- Not yet in standard ChatGPT chat, only Work, Codex and API
Best for high-volume, cost-sensitive workloads.
Gemini 4 Argon
Strong value- Index
- 53
- Price
- $2 / $10 intro
- Released
- 30 September
- Very low hallucination rate in independent tests
- Leads many agent and long-context benchmarks
- Price doubles to $4 / $20 after 30 November
Excellent for long documents and agent work; try it during the intro offer.
Meta Muse Wonder
New entrant- Index
- 52 (Full: 56)
- Price
- $4 / $20
- Released
- 12 October
- Meta's first frontier-class model
- Built around long-term memory of the user
- Full version only inside Muse Code
Interesting, but early; test before committing.
Open weights
Run it yourself- Best
- MiMo-V2.6-Pro, 46
- Also
- GLM-5.3, Kimi K3, Mistral Large 4
- Run on your own infrastructure, data never leaves
- Extremely cheap through hosted providers
- About six months behind the frontier
Ideal for private data and simple, repetitive tasks.
Which model for which job
- Best quality, cost no object
- Claude Opus 5.5 at high or max effort. Use it for complex analysis, long documents and multi-step agents.
- Best default for a team
- Claude Sonnet 5.5 or GPT-6.1 Sol. Both are near the top, at $2 / $10 per million tokens.
- Cheap and fast at scale
- Claude Haiku 5.5, GPT-6 Luna or MiMo-V2.6-Pro for classification, routing, tagging and extraction.
- Coding agents
- Claude Sonnet 5.5 and Opus 5.5 top the independent coding-agent rankings; GPT-6.1 Sol is the value pick in Codex.
- Sensitive or regulated data
- An open-weight model on your own infrastructure, or an enterprise plan with zero data retention.
- Very long documents
- All the frontier models now accept about 1M tokens; Claude has no long-context surcharge on Opus and Sonnet.
What this means for data teams
For analysts and analytics engineers, the practical lessons from this round of releases are less about who is number one and more about how to use models well:
- 1
Test on your own work
Benchmarks measure general skill. Run 20 of your real tasks, such as SQL questions against your schema, metric definitions or report summaries, through two or three models and compare the answers.
- 2
Compare cost per task, not per token
Token efficiency varies a lot. A model that is cheaper per token can still be more expensive per answer if it writes three times as much.
- 3
Route simple work to small models
Use Haiku 5.5, Luna or an open model for classification and extraction, and save the expensive models for the hard questions.
- 4
Watch the tokenizer
New model generations can count the same text differently. Haiku 5.5, for example, uses more tokens for the same text than Haiku 4.5, which trims some of its headline saving.
- 5
Give models good context
Every model answers data questions far better with clear table names, column descriptions and a semantic layer. That work pays off whichever model you choose.
Six weeks of releases
Claude Fable 5.1
Anthropic's Mythos-tier model with additional safeguards.
GPT-6 Astra
OpenAI's flagship at $10 / $50 per million tokens.
MiMo-V2.6-Pro and Grok 4.7
The strongest open-weight model, and xAI's latest.
Claude Opus 5.5, GPT-6 Sol and Luna
A new number one, plus OpenAI's mid and small tiers.
Claude Sonnet 5.5
Near-Opus quality at half the price.
GPT-6.1 Sol
Announced at DevDay 2026; the best value at the top end.
Gemini 4 Argon announced
Google's most capable model, first released to vetted cyber defenders.
Mistral Large 4 preview
1 trillion parameters, weights promised by the end of October.
Claude Haiku 5.5
The small model, about 90% cheaper per token than Haiku 4.5.
Meta Muse Wonder
Meta's first frontier-class model.
Gemini 4 Argon generally availableNow
API, Vertex AI, the Gemini app for AI Pro and Ultra subscribers.
Mistral Large 4 open weightsUpcoming
Expected release of the downloadable model.
Check your understanding
1/3Why is "cost per task" a better comparison than "price per million tokens"?
A model with a low token price can still cost more per answer if it writes far more tokens, as Grok 4.7 shows.
2/3Which model is only available inside a subscription product, not through a public API?
Muse Wonder Full is only served inside Meta's Muse Code subscription; the standard Muse Wonder is on the Meta Model API.
3/3What is the main reason to choose an open-weight model?
Open models trail the frontier, but they give you full control over where your data goes and are very cheap to run.
You got 0 out of 3.
Want to go deeper on the two biggest launches? Read GPT-6.1 Sol and OpenAI DevDay 2026 and Gemini 4 Argon, explained.
Sources: Artificial Analysis Intelligence Index and leaderboard (mid-October 2026); Anthropic API pricing; OpenAI DevDay 2026 announcements; Google DeepMind's Gemini 4 Argon announcement (30 September 2026); TechCrunch on Meta Muse Wonder (12 October 2026); Simon Willison on Mistral Large 4 (6 October 2026). Benchmark scores and prices change often; check each provider before committing.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 14 October 2026. About me