Artificial Analysis Intelligence Index v5 is the benchmark update that changes how freelancers and small teams should pick a model. Released October 27, 2026, it is Artificial Analysis’s biggest upgrade since v4 launched in January — and for the first time, 60% of the index is based on evaluations with private test sets or held-out answers, designed to stop models from training on public test questions.
The headline result is a split. Claude Opus 5.5 (max) leads at 58, ahead of OpenAI’s GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55, with Gemini 4 Argon (high) next at 50. The open-weights gap, meanwhile, widened from 3 points to 12.
What changed in v5
The index weights four categories equally at 25% each: knowledge work, coding, science, and general knowledge. New benchmarks include FrontierCode 2.0 (132 real-world coding tasks, graded on whether a maintainer would merge the change) and Terminal-Bench-Science (70 agentic research tasks).
Upgrades include HLE-Diamond, a refined 1,000-question Humanity’s Last Exam split between reasoning and knowledge. Three older benchmarks were removed. The result is harder and less gameable: median output per task rose by two-thirds, from 35,000 to 59,000 tokens.
No single model wins everything. Opus 5.5 leads three of nine evaluations; five other models lead the rest, including GPT-6 Astra on science tasks and Gemini 4 Argon on automation.
The evidence: scores, gaps, and cost per task
| Detail | What the source says |
|---|---|
| Released | October 27, 2026 |
| Private-test share | 60% of the index on private test sets or held-out answers |
| Leader | Claude Opus 5.5 (max) at 58 |
| Second / third | GPT-6 Astra (max) at 57; GPT-6.1 Sol (max) at 55 |
| Next | Gemini 4 Argon (high) at 50 |
| Top open-weights model | Kimi K3.1 Pro at 46 — weights not yet released |
| Best released open weights | DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) at 32 |
| Open-vs-closed gap | 12 points, up from 3 in v4.3; released open weights trail the leader by 26 points |
| Cheapest top-10 cost per task | GPT-6.1 Terra (max) at $0.96 per Intelligence Index task |
| Most expensive top-10 cost per task | Claude Fable 5.1 (max) at $13.04 per task |
| Opus 5.5 cost per task | $10.04 at max effort, versus $0.98 at low effort (score 58 vs 35) |
| Median output per task | 59,000 tokens, up from 35,000 in v4.3 |
Read the cost column carefully. GPT-6.1 Sol (max) scores within 4 points of Opus 5.5 (max) at about an eighth of the cost per task ($1.27 vs $10.04), and Terra (max) at $0.96 is the cheapest of the 10 highest-scoring models shown. Effort matters: Opus 5.5 scores 35 at low effort and 58 at max, at 10 times the cost.
Pick a model by the job, using v5
Bulk work — Terra or Sol — Near-frontier scores at $0.96–$1.27 per index task, not $10.
Hard coding — Opus 5.5 — Leads FrontierCode 2.0 at 46%; pay max effort only for work worth it.
Science tasks — GPT-6 Astra — Leads Terminal-Bench-Science at 63% and the physics subset at 72%.
Self-hosted — check the gap first — Released open weights top out at 32, 26 points behind the leader.
How to profit from it
Clients want the cheapest model that does their job well. Sell that decision:
- Offer a model-stack audit. Map a client’s real tasks to v5 categories — knowledge work, coding, science, general — and recommend one model per category instead of one model for everything.
- Price the savings. Show the cost-per-task gap: Terra at $0.96 versus Opus at $10.04, where the score difference does not matter. Savings you can show are savings you can charge for.
- Sell an effort-settings policy. Low effort for drafts, max effort for final deliverables. That single policy can cut a team’s model bill without touching quality where it counts.
- Warn open-weights clients honestly. If a client insists on self-hosted models, show them the 26-point gap for released weights and the science-task scores under 10% for most open models. Let them choose with eyes open.
- Re-run the audit quarterly. Artificial Analysis also launched a Progress Index tracking capability over time. A quarterly re-check is an easy retainer.
What to watch before you switch models
First, v5 leader scores use maximum effort settings. Your real cost depends on effort, prompts, and task length — test your own tasks before switching production workflows.
Second, benchmark scores are not product reviews. A coding leader may still be wrong for a client’s data rules, latency needs, or budget.
Finally, do not compare v5 scores directly with v4 scores. The index and its evaluations changed, so a model that “fell” may have faced a harder test, not gotten worse.
Source: Artificial Analysis’s official Intelligence Index v5 announcement, linked below. Scores are at stated effort settings, mostly maximum.
Frequently Asked Questions
What is the Artificial Analysis Intelligence Index v5?
A composite measure across knowledge work, coding, science, and general knowledge, released October 27, 2026, with 60% on private test sets or held-out answers.
Which model leads v5?
Claude Opus 5.5 (max) at 58, ahead of GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55.
How big is the open-weights gap?
The gap widened from 3 points in v4.3 to 12 points in v5. The best released open weights score 32, 26 points behind the leader.
What is the cheapest top model per task?
Among the 10 highest-scoring models, GPT-6.1 Terra (max) is cheapest at $0.96 per task, versus $13.04 for Claude Fable 5.1 (max).
Why did scores change so much?
Because the evaluations changed, making v5 harder and less gameable, so v5 and v4 scores are not directly comparable.
Does any model win every evaluation?
No. Six different models lead across the nine evaluations.
Official source: Artificial Analysis — Announcing Artificial Analysis Intelligence Index v5.
Keep Reading
Want more tools that actually pay for themselves? Browse the AI Tools section for the latest launches, pricing breakdowns, and profit playbooks — and check the official source linked above before you buy or deploy anything.