July 14, 2026. The launch wave of the past two weeks came with a flood of vendor reported benchmarks. The independent read has now landed: Artificial Analysis published its full evaluations of the GPT-5.6 family on July 9, alongside its Grok 4.5 analysis on July 8 and Muse Spark 1.1 on July 10. We flagged the missing third party numbers in our launch coverage, so here is the scoreboard now that it exists, with the caveats that matter.
The independent scoreboard
From Artificial Analysis's GPT-5.6 evaluation and its accompanying model analyses:
- GPT-5.6 Sol lands 1 point below Claude Fable 5 at about a third of the cost. Sol on max effort scores 59 on the Intelligence Index at 1.04 dollars per task. Terra scores 55 at 0.55 dollars and Luna 51 at 0.21 dollars per task.
- Sol leads the new Coding Agent Index at 80, topping all three evaluations in OpenAI's Codex harness and tying Grok 4.5 on one, with per task costs around 40 percent below Claude Fable 5 and 10 percent below Opus 4.8 in Claude Code.
- The skip Terra finding. Across effort levels, Luna and Sol sit on the price to intelligence frontier while Terra never does: for any Terra configuration there is a Luna or Sol setting that is smarter for the same money or equal for less.
- Grok 4.5 scores 54, fourth overall behind Fable 5, GPT-5.5, and Opus 4.8, at 2 dollars per million input tokens, confirming the cheap frontier tier we covered in yesterday's Grok 4.5 brief.
- Muse Spark 1.1 scores 51, an 8 point gain for Meta in three months, arriving alongside Meta's first paid model API.
- The catch: hallucination rate went up. On the Omniscience evaluation, Sol posts a small accuracy uplift over GPT-5.5 paired with a higher hallucination rate, and Fable 5 stays clearly ahead on analytical rubric scores in the knowledge work benchmark.
What it means for operators
First, route on independent cost per task, not sticker price or vendor slides. The gap between 1.04 and 0.21 dollars per task across one model family is a budget decision hiding in a dropdown. Second, act on the Terra finding: if your workflows defaulted to the middle tier after the GA week, re-run the math against Luna and Sol effort levels. Third, a higher hallucination rate on a cheaper, smarter model is exactly why verification layers exist: keep machine checkable outputs and adversarial review in your automations, the pattern from the 64 agent proof playbook. Model choice is now a quarterly routing exercise, and building automation that stays model agnostic is the hedge. If that math is not something your team wants to own, it is a solved problem for a dedicated AI engineer.
Frequently Asked Questions
It is an independent composite benchmark run by Artificial Analysis across reasoning, coding, math, and knowledge evaluations, with published cost per task figures. Because the same harness runs every model, it is one of the few places to compare vendor models without relying on the vendors own reported numbers.
On the Intelligence Index, Sol on max effort scores 59 versus Fable 5 at 60, at roughly one third of the cost per task. But Fable 5 keeps a clear lead on analytical quality in the knowledge work benchmark, and Sol shows a higher hallucination rate than its predecessor, so the right answer depends on the task and your verification setup.
The independent data suggests skipping Terra: Luna and Sol define the cost to intelligence frontier, so for any Terra setting there is a Luna or Sol configuration that is equal or better value. Route cheap, high volume work to Luna effort levels and reserve Sol max for tasks where the outcome justifies about a dollar per task.
Vendor benchmarks are selected to flatter the launch. Independent evaluations run identical harnesses across competitors and publish cost per task, which is the number that actually drives your AI budget. The July wave showed the pattern again: the marketing and the independent scoreboard agree on headlines but differ on the caveats, like hallucination rates.