Best AI Model for Coding and Agents: November 2026 Comparison
Gemini 4 Carbon, Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol compared on independent data. Why effort settings matter more than the model, and what to pick for each job.
Asking which AI model is best has stopped being a useful question. The top models now sit within a few points of each other on the leading index, while their cost per task differs by more than ten times. The decision that actually moves your product is which model, at which effort level, for which job.
Here is where things stand in November 2026, using independent measurements wherever they exist and labeling vendor numbers as vendor numbers. These are published results, not my own benchmark run.
The current leaderboard
These figures come from the Artificial Analysis leaderboard, read on November 5, 2026. It is an independent test of general intelligence, not coding specifically, and each row shows the model at its highest-scoring effort setting. Scores update without notice, so check the live page before you rely on one.
| Model (effort) | Index | Cost per task | Time per task |
|---|---|---|---|
| Gemini 4 Carbon (high) | 60 | $7.76 | 30.3 min |
| Claude Opus 5.5 (max) | 59 | $10.04 | 26.2 min |
| GPT-6 Astra (max) | 57 | $6.24 | 19.0 min |
| GPT-6.1 Sol (max) | 55 | $1.27 | 21.2 min |
| Claude Sonnet 5.5 (max) | 55 | $10.02 | 30.7 min |
| GPT-6.1 Terra (max) | 51 | $0.96 | 13.1 min |
| Gemini 4 Argon (high) | 50 | $9.95 | 26.2 min |
| Claude Fable 5.1 (xhigh) | 50 | $10.85 | 18.6 min |
One catch sits at the top of this table. Carbon is not generally available yet. Google's October 28 announcement says it is going to a subset of cyber defenders first, with developers to follow "as soon as possible" and no date. If you are shipping this quarter, plan around the other rows.
Effort matters more than the model
The most useful data on that leaderboard is further down, where each model appears at several effort levels. Compare these results:
| Model (effort) | Index | Cost per task | First token |
|---|---|---|---|
| GPT-6 Astra (medium) | 52 | $2.73 | 6.5s |
| GPT-6.1 Sol (high) | 50 | $0.59 | 37.8s |
| Gemini 4 Argon (high) | 50 | $9.95 | 40.4s |
| Claude Opus 5.5 (medium) | 48 | $2.45 | 18.3s |
| GPT-6.1 Sol (medium) | 46 | $0.36 | 6.3s |
| GPT-6.1 Terra (high) | 42 | $0.26 | 10.7s |
Two models at the same score of 50 cost $0.59 and $9.95 per task. Meanwhile, running Opus 5.5 at max instead of medium adds 11 index points but multiplies cost by four and takes time to first answer token from 18 seconds to 343. Max effort is a setting for overnight agents, not a sensible default for a product with users waiting.
What it costs
| Model | Input per 1M | Output per 1M | Note |
|---|---|---|---|
| GPT-6 Astra | $10 | $50 | $20 / $75 over 272K input |
| Claude Fable 5.1 | $10 | $50 | |
| Claude Opus 5.5 | $4 | $20 | |
| Gemini 4 Argon | $2 | $10 | $4 / $20 from Dec 1 |
| Gemini 4 Carbon | $2 | $10 | Intro, then $4 / $20 |
| Claude Sonnet 5.5 | $2 | $10 | |
| GPT-6.1 Sol | $2 | $10 | $4 / $15 over 272K input |
| GPT-6.1 Terra | $1 | $5 | $2 / $7.50 over 272K input |
Prices from OpenAI, Anthropic and Google, plus Google's Carbon announcement. List price is only half the story, because models use very different token counts. Sonnet 5.5 and Sol share a list price, yet at max effort Sonnet used 335M output tokens on the index run against Sol's 63.7M. My Gemini 4 Argon guide covers the December price change in detail.
For coding specifically
This is where independent data gets thin. Artificial Analysis's Coding Agent Index reports Claude Code with Opus 5.5 (max) at 62 and Antigravity CLI with Carbon (high) at 59. Those are the only two scores I could find on it, and it measures a model and a tool together.
Beyond that, you are mostly reading vendor tables. Anthropic's own table puts Opus 5.5 ahead of GPT-6 Astra on FrontierCode v1.1 (54.4% to 53.3%), a narrow gap, and behind it on AutomationBench and Terminal-Bench-Science. Terminal-Bench 4.0 results vary by who reports them and at what effort, so I would not quote a single winner from it. Be skeptical of any article that declares one, this one included.
What I would pick for each job
- Long autonomous coding runs: Claude Opus 5.5, which has the top coding agent score available today. Watch Carbon once developers get access.
- Default for product features: GPT-6.1 Sol at medium or high. Its cost per task is hard to beat at this quality level.
- High-volume simple work: GPT-6.1 Terra at low or medium, with a first token in the low single digits of seconds.
- Very long context: test Gemini 4 Argon. Google's model card shows a strong lead on long-context reasoning, though that is a vendor result.
- Hardest reasoning, cost secondary: GPT-6 Astra or Opus 5.5 at high effort, on a case-by-case basis.
How to decide for your product
- Build a small eval from 20 to 50 real tasks, including your failures.
- Run two or three models at two effort levels each.
- Record pass rate, cost per successful task, and time to first token.
- Send each type of task to the cheapest setup that passes, and rerun the eval whenever a model changes.
The leaderboard will look different next month. An eval and a routing layer keep working when it does. Once you are routing, cost control becomes the next problem to solve.
Need help choosing and wiring up models for a real product? I take on remote AI product work for startups in the US and UK.
