Claude Opus 5.5 vs Grok 4.7: Two Flagships, Twenty-Four Hours Apart

Grok 4.7 arrived on September 21, 2026. Claude Opus 5.5 arrived the next day. Two frontier flagships inside twenty-four hours, aimed at the same buyer from opposite directions: Grok 4.7 held its predecessor's price at $2 per million input tokens and $6 output, while Opus 5.5 cut its predecessor's from $5/$25 to $4/$20. On the sticker, Grok 4.7 looks like half the cost of Claude Opus 5.5 and then some. It is not, and the reason is the most useful thing in this comparison. Grok 4.7's rates double to $4/$12 once a prompt passes 200,000 tokens, and the model is markedly verbose — Artificial Analysis measured a 113% increase in reasoning-token use over Grok 4.6, with roughly 81,000 output tokens consumed per agentic task. A cheaper rate on a model that writes more is not a cheaper task, and the two effects have to be multiplied before either sticker means anything.

What each vendor actually shipped

SpaceXAI released Grok 4.7 on Monday, September 21 — the same $2/$6 and the same 500,000-token context window as Grok 4.6, with a later knowledge cutoff and a new top reasoning level. The vendor reports it built on a larger base model than its predecessor and positions it for coding, agentic work and knowledge labour; it ships live in Cursor and in the company's own Grok Build terminal agent. The company's own Terminal-Bench 4.0 figure is 38%, and Artificial Analysis's independent run of the same benchmark put it at 33% — a five-point disagreement between vendor and evaluator on the same axis, which is worth remembering when you read any single number from either side.

Anthropic released Claude Opus 5.5 on Tuesday, September 22 at $4/$20, down from Opus 5's $5/$25, with cache reads falling from $0.50 to $0.20 per million and the vendor claiming typical workloads cost about 40% less to run than its predecessor. It keeps the 1M-token context and 128K output ceiling of Opus 5 and adds a Fast mode at $8/$40 that Anthropic describes as a research preview. The capability claims are vendor-run against Anthropic's chosen competitors: Terminal-Bench 4.0 at 66.4%, FrontierCode v1.1 at 54.4%, GDPval-AA v2.1 at 1,846 Elo.

Set those two paragraphs side by side and the shape of the matchup is already visible. One vendor cut the price of a model that improved. The other held the price of a model that improved. Both are making a bet about where the market's floor is, and they disagree.

The rate card, and the two rows that change it

Figures below are from OrcaRouter's catalogue, read on September 24, 2026, passed through at 0% markup — the provider list price with no spread added.

• Input — Claude Opus 5.5 at $4.00 per million tokens vs Grok 4.7 at $2.00; Grok is 50% cheaper

• Output — $20.00 per million vs $6.00; Grok is 70% cheaper

• Cache read — $0.20 per million vs $0.50; the only pricing line where Opus 5.5 is cheaper, and it is 60% cheaper

• Context — 1M tokens for Claude Opus 5.5 vs 500,000 for Grok 4.7, exactly double

• Maximum output — 128,000 tokens for Claude Opus 5.5 vs 450,000 for Grok 4.7, more than three times as much

• Modality — both accept text, images and files and return text

The first modifier is the tier step. Grok 4.7's $2/$6 applies below a 200,000-token prompt; above it, both rates double to $4/$12. That puts Grok's long-prompt input rate exactly level with Claude Opus 5.5's flat $4.00 and narrows the output advantage from 70% to 40%. On the very long prompts where Grok's 450,000-token output ceiling is the reason you would choose it, you are paying Opus 5.5 input rates to get there.

The second modifier is verbosity, and it is larger. Artificial Analysis measured Grok 4.7 using 113% more reasoning tokens than Grok 4.6 on the same evaluations, with around 81,000 output tokens consumed per agentic task in one published account. Reasoning tokens bill as output tokens on every API in this comparison. At Opus 5.5's $20.00 per million, 81,000 output tokens costs $1.62. At Grok 4.7's $6.00, the same 81,000 costs $0.49 — still a third of the price, so the advantage survives — but the ratio is now 3.3× rather than 3.3× on paper and something considerably smaller in practice once Opus 5.5's lower cache-read rate is applied to the re-sent prefix of a long agent session. The 70% output saving advertised on the rate card is not the saving you will book.

If you want the single clearest illustration of why token price is the wrong unit: the same 30,000-input, 8,000-output agentic request costs roughly 28 cents on Opus 5.5 and 11 cents on Grok 4.7. Make the input 250,000 tokens instead and Grok's doubled rate brings its cost to about $1.10 against Opus 5.5's $1.16. The sticker gap went from 2.5× to a rounding error, with no change except the length of the prompt.

The independent scores point one way

On Artificial Analysis's Intelligence Index, Claude Opus 5.5 sits at 58 in its maximum-effort configuration and 54 at high effort, at the top of the models that evaluator tracks. Grok 4.7 scores 46 — two points above Grok 4.6, and twelve below Opus 5.5's maximum-effort reading. That is the widest gap in this piece and it is third-party, not vendor.

The same evaluator's other boards are more flattering to Grok, and the spread between them is the interesting part. On the Coding Agent Index, Grok 4.7 scores 56 against Grok 4.6's 47 — a nine-point gain, larger than its two-point gain on general intelligence, which suggests the release is a coding update more than a general one. On AA-Briefcase, its long-horizon knowledge-work board, Grok 4.7 lands at 1,657 Elo in fourth place, behind three Anthropic entries and ahead of Claude Opus 5 at xhigh effort. Three boards, three different verdicts, all from the same evaluator on the same day.

Then there is the one axis both sides measure: Terminal-Bench 4.0. Anthropic reports 66.4% for Opus 5.5 at xhigh effort. SpaceXAI reports 38% for Grok 4.7, and Artificial Analysis's independent run put it at 33%. Whichever number you take for Grok, the gap to Opus 5.5 is between 23 and 33 points. Terminal-Bench is a single harness and a vendor-run one on the Anthropic side, so it is not a verdict on either model — but it is not a close call on the axis, and it is the axis most closely tied to the agentic coding work both vendors are selling into.

Where each one is the right answer

Grok 4.7's advantage is real and structural in two places. The first is short-prompt, high-volume work: below 200,000 tokens with modest output, $2/$6 against $4/$20 is a genuine 2.5× to 3.3× difference and the verbosity penalty is a smaller absolute number when the whole task is small. The second is output ceiling — 450,000 tokens is more than three times Opus 5.5's 128,000, and for bulk generation of long artifacts that is the constraint that decides whether a job is possible at all.

Claude Opus 5.5's advantage is in the agentic long run. Twice the context window at 1M tokens, a cache-read rate 60% lower, and the strongest independent intelligence score of the pair. Every one of those compounds with session length: the cache read is paid on the re-sent prefix on every turn, so a 60% cut on that line does more for a two-hour agent session than a 70% cut on output does for a one-shot generation.

One operational note that belongs with any Opus 5.5 decision: the model routes most cybersecurity requests to Claude Opus 4.8 and biology work to Claude Opus 5 unless your account is verified through Anthropic's Life Sciences or Cyber Verification programmes. In those domains the model string in your logs is not always the model that answered. Grok 4.7 has no equivalent rerouting, though its own model card reportedly shows 3.3% of risky requests still passing in high mode with slightly higher rule-violation rates than Grok 4.6 — a vendor-published safety figure on the other side of the same ledger.

Both models are on OrcaRouter at provider list price with 0% markup — Claude Opus 5.5 at $4.00/$20.00 with $0.20 cache reads, Grok 4.7 at $2.00/$6.00 with the >200K tier applied as the provider sets it — behind one key, so the split above can be measured on your own traffic rather than argued about. Where both are routed, automatic failover sits under either, which matters more than usual for two models released a day apart and still settling.

The verdict this matchup earns: Grok 4.7 is the cheaper model on most workloads and the one to pick for long outputs; Claude Opus 5.5 is the stronger one where sessions are long, context is deep and the work is agentic. The 24-hour launch window made them look like rivals. The rate cards and the token counts make them a split.