On October 27, 2026, Artificial Analysis (AA) shipped Intelligence Index v5, its largest methodology overhaul since v4 launched in January. The new top three: Claude Opus 5.5 at 58, GPT-6 Astra at 57, and GPT-6.1 Sol at 55.
Alongside the scores, AA published measured per-task costs. At max reasoning effort, Opus 5.5 runs $10.04 per task. Sol runs $1.27. AA's own summary line: Sol scores within 4 points of the leader at roughly an eighth of the per-task cost.
Here is the short version:
- Scores: Opus 5.5 leads at 58, Sol sits third at 55 — a 3-point gap inside the same tier.
- Bills: $10.04 versus $1.27 per task at max effort — almost an order of magnitude apart.
- Why Sol is cheaper: not token pricing. Opus 5.5 produces roughly 3x the output tokens per task, and 2x unit price times 3x usage is the 8x bill.
- How to choose: run everyday agent workloads on Sol (or Terra), and reserve Opus 5.5 at max effort for the hardest problems.
Why v5 is a big deal
Two v5 changes matter most for practitioners. First, tasks got heavier: median benchmark output grew from 35K tokens under v4 to 59K, with roughly 90% of that output coming from agentic multi-step work, so the index tracks real workloads more closely. Second, contamination control: 60% of the evaluations moved to private test sets, closing the benchmark-gaming lane.
Two more data points from the leaderboard: GPT-6 Astra slots in second at 57, and Kimi K3.1 Pro takes the top open-weights spot at 46 — but the open-versus-closed gap at the top widened from 3 to 12 points.
One clarification that often gets lost: Opus 5.5 also topped AA's chart back in September, but that was the older v4.3 task set. Holding the #1 spot after a full methodology revision is a stronger claim than the original one, and the two events should not be conflated.
Scores and bills in one table
Per-task figures below are AA's official v5 announcement numbers, measured at max effort:
| Model | v5 score (max) | Cost per task (AA measured) |
|---|---|---|
| Claude Opus 5.5 | 58 | $10.04 |
| GPT-6 Astra | 57 | $6.24 |
| GPT-6.1 Sol | 55 | $1.27 |
| GPT-6.1 Terra | outside top 5 | $0.96 (cheapest) |
| Claude Fable 5.1 | outside top 5 | $13.04 (priciest) |
$10.04 divided by $1.27 is about 7.9, matching AA's "roughly an eighth" framing. AA also notes that per-task costs across the frontier span more than 13x. Inside the narrow "flagship capability" band, the bills diverge far more violently than the scores.
A caution if you read third-party coverage: you may see Sol at $0.72 and Opus 5.5 at $5.98 per task. Those are legacy numbers from the v4 task set, where median output was only 35K tokens. With v5's heavier tasks, the official figures are $1.27 and $10.04. Cite the October 27 announcement.
Why per-token pricing cannot predict per-task cost
This has been the most debated point since v5 dropped. At the official list rates AA prices against, Opus 5.5 costs about 2x more per token than Sol (HiAPI's platform rates differ — see below), yet lands 8x more expensive per task. The missing multiplier is output length: measured runs show Opus 5.5 generating roughly 3x more output tokens per task. Two times unit price multiplied by three times usage is the eight-times bill.
v5 makes this worse, because heavier tasks mean longer outputs, and models that write more get more expensive faster. It also explains why reading price lists systematically misleads: a price list tells you what a token costs, an invoice tells you what a finished job costs. Both matter, but only one belongs in a budget.
How the index is built, and why effort settings matter
The v5 index combines nine evaluations across four equal categories, 25% each: knowledge work, coding, science, and general capability. No model sweeps all nine — Opus 5.5 leads three, GPT-6 Astra leads two.
The other notable v5 shift is how much effort settings now move the scores. Opus 5.5 scores 35 at low effort versus 58 at max; Sol's effort spread widened to 20 points. In other words, those 58 and 55 leaderboard numbers are full-throttle results.
The practical read: scores are produced at max effort, and the invoices are burned at max effort too. When you budget from a leaderboard, always read the score and the cost together with the effort tier. At the low or medium settings everyday work actually uses, the real gap between the two models looks different from the table.
Route by task difficulty instead of picking a side
Given this data, the cheapest setup is neither "all Sol" nor "all Opus" — it is a difficulty ladder:
- Everyday agent work (scripts, data cleaning, routine coding, batch jobs): Sol, or even Terra at $0.96 per task, carries the volume.
- Hard problems (complex refactors, deep reasoning, multi-hop analysis): spend on Opus 5.5 at max effort, where the 3-point advantage actually shows.
- One more cost lever: adjust effort settings per request instead of locking everything to the top tier.
This "one task pool, two models" routing pattern is exactly where an API aggregator earns its keep — one key, one bill, and per-request model switching. We walk through the same task-based switching approach for Claude Code and Codex in a companion guide if you want a concrete setup.
Using both models on HiAPI
Platform pricing first (as of October 2026, base tier up to 272K context):
| Model | Input | Output | Cache read |
|---|---|---|---|
| claude-opus-5-5 | $2.02/M | $10.08/M | $0.10/M |
| gpt-6.1-sol | $0.50/M | $2.50/M | $0.025/M |
Keep the two number systems separate: the table above is HiAPI per-token pricing, while the $10.04 / $1.27 figures earlier are AA's measured per-task costs. On our price list, Sol's input and output rates are each about a quarter of Opus 5.5's — that is the rate-card view. Fold in the ~3x output-volume difference described above and you arrive at AA's one-eighth per-task view. Both are correct; do not mix them.
On the integration side, Opus 5.5 speaks both Anthropic Messages and OpenAI Chat Completions, with five effort tiers to tune. Sol supports Responses and Chat Completions with a 1.05M-token context window. One API key reaches both, and routing is a per-request choice.
- Full capability breakdowns: Claude Opus 5.5 model page and GPT-6.1 Sol model page
- Hands-on guides: How to use GPT-6.1 Sol on HiAPI and Claude Opus 5.5 API pricing
- Earlier comparison for context: Claude Opus 5.5 vs GPT-5.6 Sol
- Same-batch coverage: GPT-6.1 Terra price cut, Switching models in Claude Code and Codex on HiAPI, and Genie 4 world model
FAQ
How is Sol's $1.27 per-task cost measured?
It is AA's own measured figure from the v5 announcement: run a suite of tasks at max effort, then price the actual input and output tokens consumed at list rates. It is an invoice, not a sticker — which is why it tracks what you actually pay more closely than any price list.
Which is right: Sol at $0.72 or at $1.27 per task?
$0.72 (and $5.98 for Opus 5.5) are legacy figures from the v4 task set, where median output was only 35K tokens. With v5's heavier tasks the official numbers moved to $1.27 and $10.04. Cite the October 27, 2026 v5 announcement.
How many times cheaper is Sol per token?
On HiAPI pricing as of October 2026 (base tier up to 272K context), Sol is $0.50/M input and $2.50/M output versus $2.02/M and $10.08/M for Opus 5.5 — about 4x on rates. That is the rate-card view. AA's per-task view is ~8x because Opus 5.5 also generates about 3x more output tokens per task. Both views are correct; just do not blend them.
What is the effort setting, and why do scores swing so much?
Effort is the model's reasoning-intensity tier: higher means deeper thinking and longer outputs. Under v5, Opus 5.5 spans 35 to 58 across tiers and Sol's spread reaches 20 points. So whether Sol can replace Opus 5.5 depends on which tier your everyday work runs at — and how much you are willing to spend when it matters.
Should I use Sol or Opus 5.5 for my workload?
Start everyday agent work (scripts, cleaning, routine coding) on Sol, where volume is affordable. For hard problems, compare a batch on Opus 5.5 at max effort. One HiAPI key reaches both models, so testing the split costs nothing extra in integration work.
Bottom line
The v5 leaderboard puts two facts side by side. On capability, Opus 5.5, Astra, and Sol are compressed within 3 points — nobody dominates. On cost, per-task bills diverge by 8x, far more than the scores do. Scores are made at max effort, and so are the invoices, so always read both numbers with their effort tier attached. For most teams the right move is not picking a side but routing by difficulty: Sol or Terra for volume, Opus 5.5 at max effort for the hardest tasks. Both models are available on HiAPI today, ready for that routing setup.
