Three frontier or near-frontier models shipped within 48 hours of each other in September 2026 — Grok 4.7 on September 21, then Claude Opus 5.5 and GPT-6 Sol both on September 22 — and each company published its own benchmark suite comparing itself against a different, self-selected set of competitors. There is no single third-party table putting all three head-to-head. This post pulls together the one benchmark all three companies' own materials actually reported a number for, plus pricing, and is honest about where the comparison stops being possible with public data.
TL;DR
| Model | Terminal-Bench 4.0 | Input price | Output price |
|---|---|---|---|
| Claude Opus 5.5 (xhigh) | 66.4% | $4/M | $20/M |
| GPT-6 Sol | 43% | $2/M | $10/M |
| Grok 4.7 | 38.0% | $2/M | $6/M |
Source note: Opus 5.5's figure is Anthropic's own reported result. GPT-6 Sol's is Artificial Analysis' independent measurement, not OpenAI's own launch materials (which didn't include a directly comparable Terminal-Bench 4.0 number for Sol specifically). Grok 4.7's is SpaceXAI's own published table.
The only clean comparison available
Terminal-Bench 4.0 tests whether an agent can complete complex, multi-step tasks inside a command-line interface — a reasonable proxy for real agentic coding work. It's genuinely useful here precisely because all three companies' own disclosures, directly or via independent evaluators, include a number on it. On that one shared axis, the gap between Opus 5.5 and the other two is not subtle: 66.4% against 43% and 38.0% is a 23-to-28 point spread, not a close call decided by methodology quibbles.
What each company actually chose to compare against
This is the more revealing part of the story than the raw numbers. None of the three companies benchmarked itself directly against both of the other two — each picked its own reference point:
- Anthropic's Opus 5.5 launch compared primarily against Claude Fable 5.1, Claude Opus 5, GPT-6 Astra, and GPT-5.6 Sol — not GPT-6 Sol (which hadn't launched yet at the time Anthropic's benchmark page was built) and not Grok 4.7 at all.
- OpenAI's Sol/Luna launch compared against GPT-6 Astra, GPT-5.6 Sol, Claude Opus 5, and Claude Fable 5.1 — again, no Grok comparison, and Opus 5.5 hadn't shipped yet either.
- SpaceXAI's Grok 4.7 launch compared against Grok 4.6, Claude Fable 5.1, and GPT-5.6 Sol Max — its own table shows Fable 5.1 leading CursorBench 4.0 (51.8% vs Grok's 46.3%) and Terminal-Bench 4.0 (57.9% vs 38.0%), with Grok 4.7 winning specifically on EEBench (19.6% vs Fable's 6.7%) and the Harvey Legal Agent Benchmark.
Every company's benchmark page is, unsurprisingly, built to make its own model look strongest against the comparisons it chose. That's not dishonest by itself, but it does mean a genuinely fair three-way comparison requires pulling from multiple sources rather than trusting any single company's page — which is exactly the exercise this post is doing.
Pricing tells the opposite story from performance
Grok 4.7 is the cheapest of the three at $2/M input and $6/M output — identical to its Grok 4.6 predecessor's pricing, per SpaceXAI's "same price, better model" framing. GPT-6 Sol sits close behind at $2/M input and $10/M output. Opus 5.5 is meaningfully more expensive on both ends, at $4/M input and $20/M output — roughly double GPT-6 Sol's rate and more than triple Grok 4.7's output price specifically.
That inverse relationship — cheapest model scoring lowest on the one shared benchmark, most expensive scoring highest — is a normal market pattern, not an anomaly, but it does mean cost-per-completed-task, not cost-per-token, is the number that actually matters for a real decision. None of the three companies published that figure in a way that lets a reader compute it across all three models on the same task, which is the central limitation of this whole comparison.
Why Grok 4.7 isn't really positioned against the other two
It's worth being direct about this: Grok 4.7's own launch materials don't frame it as competing with Opus 5.5 or GPT-6 Sol at all — SpaceXAI's comparison table benchmarks it against Grok 4.6 and Fable 5.1, treating it as an iterative, same-price update rather than a new flagship tier. Anecdotal reaction backs that positioning — a viral X thread cited in explainx.ai's own Grok 4.7 coverage found it noticeably weaker than Sonnet 5 on a hard Three.js frontend task, with heavier token burn than a comparable Kimi run. Grok 4.7 belongs in this comparison because of its release timing and its genuine Terminal-Bench 4.0 number, not because SpaceXAI is claiming it beats either Anthropic or OpenAI's latest releases.
What each company's naming and versioning tells you about intent
There's a useful signal buried in how each company version-numbered its release this cycle, separate from the benchmark numbers themselves. Anthropic jumped straight to "Opus 5.5" rather than a smaller point release, a naming choice that reads as a deliberate signal of a meaningful, not incremental, capability jump — consistent with the size of its Terminal-Bench 4.0 lead over the field. OpenAI's "GPT-6 Sol" sits explicitly one tier below its own flagship "GPT-6 Astra," a naming structure that telegraphs cost-optimization as the point rather than raw capability, which matches its lower Terminal-Bench 4.0 score relative to Opus 5.5. SpaceXAI's "Grok 4.7" is the most conservative naming of the three — a minor point release above Grok 4.6, not a new major version — which lines up with SpaceXAI's own framing of it as a same-price iterative update rather than a capability leap, and with it posting the lowest score of the three on the one shared benchmark.
That consistency between naming convention and actual benchmark positioning is worth noting specifically because it means each company's naming scheme is a reasonably reliable, if informal, signal of how it wants a release to be read relative to its own prior generation — useful shorthand when deciding how much attention a given release deserves before digging into the full benchmark tables.
The practical question this raises: how many models should you actually track?
Three major or near-major releases within 48 hours is a genuinely hard pace to keep up with as a builder, and it's worth asking directly whether tracking all three in real time is actually the right use of attention versus checking in periodically once each ecosystem has had time to mature. The honest answer, based on the pattern visible across this comparison: Opus 5.5's lead is large enough on the one shared benchmark that it's worth knowing about immediately if agentic coding performance matters to your work, while GPT-6 Sol and Grok 4.7's changes are genuinely incremental relative to their own predecessors — useful to know exists, less urgent to switch to on release day specifically. A reasonable default: track flagship-tier releases (Astra, Fable, Opus, Grok's actual next major version) closely, and treat mid-tier or point releases (Sol, Luna, incremental Grok updates) as worth a glance at pricing and one headline benchmark rather than a full re-evaluation of your current setup.
Honest limitations
- This is not a like-for-like, independently reproduced benchmark — it stitches together numbers from three different companies' self-reported results plus one independent evaluator (Artificial Analysis), on a single shared benchmark, not a controlled head-to-head run by one neutral party.
- Terminal-Bench 4.0 measures one specific skill (command-line agentic task completion) and should not be read as a general intelligence ranking — each model may lead or lag differently on reasoning, knowledge work, or creative tasks, none of which this post attempts to compare across all three.
- Cost-per-completed-task, the number that actually determines real-world spend, isn't available across all three models from public sources — the pricing comparison here is strictly per-token list price, which doesn't account for how many tokens each model typically uses to finish a comparable task.
- GPT-6 Sol's Terminal-Bench 4.0 figure comes from Artificial Analysis, not OpenAI's own launch page — OpenAI's own materials didn't include a directly stated Sol-specific Terminal-Bench 4.0 score at launch.
What this means for builders
If Terminal-Bench-style agentic coding performance is the deciding factor and budget allows it, Opus 5.5's lead on the one benchmark all three companies engage with is large enough to matter, not a rounding error. If cost per token is the binding constraint and the task doesn't require frontier-level agentic reasoning, GPT-6 Sol and Grok 4.7 are both credible, meaningfully cheaper options — but neither has published data letting you predict cost-per-completed-task against Opus 5.5 directly, so budgeting off list price alone risks being wrong in either direction once actual token usage per task is accounted for.
Related on explainx.ai
- Claude Opus 5.5 Launch: Every Benchmark and Reaction
- GPT-6 Sol and Luna Launch: 50% Price Cuts and Where They Actually Land
- Grok 4.7 Launch: Benchmarks and Pricing
- GPT-6 Sol vs Claude Opus 5.5: What Actually Overlaps
- How to Read AI Benchmarks Without Getting Fooled
Primary sources: Anthropic's Opus 5.5 announcement, OpenAI's GPT-6 Sol and Luna announcement, SpaceXAI's Grok 4.7 announcement, and Artificial Analysis independent benchmarks, all September 21-22, 2026.
This post compares publicly disclosed benchmark and pricing figures as of September 23, 2026. Figures are subject to revision by each company and by independent evaluators.