Grok 4.7 vs Claude Opus 5.5: Benchmarks, Pricing, and Cost per Task (November 2026)

	Grok 4.7	Claude Opus 5.5
Vendor	xAI	Anthropic
Released	September 21, 2026	September 22, 2026
Input per 1M tokens	$2 ($4 once a prompt reaches 200K tokens)	$4
Output per 1M tokens	$6 ($12 once a prompt reaches 200K tokens)	$20
Cached input read	$0.50	$0.20
Context window	500K	1M
Max output	Not listed	128K
AA Intelligence Index	46	59 (max effort)

Grok 4.7 and Claude Opus 5.5 shipped one day apart in September 2026, and they answer two different questions. xAI's pitch is a frontier class model at a price that used to buy a mid tier one. Anthropic's pitch is its recommended default for everyday coding and knowledge work. On paper Grok is the bargain. In measured runs the answer is more complicated, and the complication is the most useful thing in this comparison.

Benchmarks: Opus wins the coding rows

Benchmark	Grok 4.7	Opus 5.5	Source
Terminal Bench 4.0	37.6%	66.4%	Each vendor's own run
CursorBench 4.0	46.3%	57.8%	Each vendor's own run
SWE Together (pass at 1)	64.7%	68.8%	Independent
DeepSWE v1.1	71.0%	Not published	xAI
Harvey Legal Agent Benchmark	19.6%	3.8%	xAI's sheet
AA Intelligence Index	46	59	Artificial Analysis

Every shared coding row goes to Opus 5.5. The Terminal Bench 4.0 gap is close to 29 points, which in a terminal agent loop is the difference between a model that finishes multi step tasks and one that stalls on them. CursorBench 4.0 is closer at 11.5 points, and the one independent row, SWE Together, is closer still at about 4 points. That pattern is worth noticing: the gap narrows on the benchmark nobody's marketing team ran.

Grok 4.7 has two results worth taking seriously. DeepSWE v1.1 at 71.0% is a strong number, though Anthropic has not published an Opus score on that test, so there is nothing to compare it with. And on Harvey's Legal Agent Benchmark Grok posts 19.6% against 3.8% for Opus. That figure comes from xAI's own sheet, so it wants an independent rerun before a legal team builds on it, but a fivefold difference is not noise. We covered what xAI changed in training to get here in our Grok 4.7 benchmark breakdown.

The price list says Grok. The invoices say it depends.

Per token, this is not a contest. Grok 4.7 is half the input price of Opus 5.5 and 30% of the output price.

Per task, the picture inverts. In Artificial Analysis's index runs, Grok 4.7 at its default high effort cost about $2.73 per task. Opus 5.5 at its default medium effort cost about $1.34 per task on roughly 26,000 output tokens. The cheaper model cost twice as much to finish the same work.

xAI's own data explains it. Grok 4.7 averages about 70,000 output tokens per CursorBench task at its highest effort setting, against about 15,700 at low. Reasoning tokens bill as output, so a model that thinks at length erases its own rate card advantage. On SWE Together, Grok's measured cost was $7.81 per task.

Three structural details push the same direction:

1. Long prompts double Grok's price. Once a prompt reaches 200K tokens, xAI bills the whole request at $4 input and $12 output. At that point the input price matches Opus exactly.
2. Cache reads favor Opus. Opus 5.5 reads cached input at $0.20 per million tokens against $0.50 for Grok. Agent loops that resend a large stable prefix on every turn are mostly cache reads, so this line matters more than the headline input price.
3. Opus has twice the context. 1M tokens against 500K, with no long context tier.

A quick illustration using list prices. A request with 50K input tokens and 10K output tokens costs about $0.16 on Grok 4.7 and $0.40 on Opus 5.5. Grok wins comfortably. Make that a 250K token prompt with the same output and Grok lands at about $1.12 against $1.20 for Opus, nearly a tie before caching, and behind once the prefix is cached. The crossover is real and it arrives at prompt sizes that agentic coding hits routinely.

Where each one belongs

Choose Grok 4.7 when the work is high volume, the prompts are short, and the task is well inside the model's ability: classification, extraction, summarization, routine code edits, first pass drafting. Pin the effort setting to low or medium and measure output tokens. At 82.6 tokens per second at its highest effort setting, it is also quick. If the legal benchmark result holds up under independent testing, legal document workflows become a second niche.

Choose Opus 5.5 when the work is agentic coding, terminal automation, or anything with long context and a large cached prefix. It has the higher success rate on every shared coding benchmark and the lower measured cost per task, which is a rare combination. If you are deciding between Opus and Anthropic's more expensive tier instead, see Fable 5.2 vs Opus 5.5.

Run both when you have a routing layer. The two are complements more than rivals: Grok for the cheap, simple majority of requests and Opus for the hard minority is a sensible split, provided the router is driven by measured task difficulty rather than guesswork. Google's Gemini 4 Carbon now competes for both slots, which we cover in Carbon against the field.

FAQ

Is Grok 4.7 cheaper than Claude Opus 5.5? Per token, yes: $2 and $6 against $4 and $20. Per completed task, not reliably. Independent runs measured Grok 4.7 at about $2.73 per task at its default effort and Opus 5.5 at about $1.34, because Grok generates far more output tokens.

Which is better for coding? Opus 5.5 leads every shared coding benchmark: 66.4% against 37.6% on Terminal Bench 4.0, 57.8% against 46.3% on CursorBench 4.0, and 68.8% against 64.7% on the independent SWE Together test.

Does Grok 4.7 beat Opus 5.5 anywhere? On price per token, on speed, and on Harvey's Legal Agent Benchmark (19.6% against 3.8%), though that legal result is from xAI's own sheet and has not been independently reproduced.

What happens to Grok 4.7 pricing on long prompts? Once a prompt reaches 200K tokens the entire request bills at $4 input and $12 output, double the standard rate.

When this applies to your stack

The lesson from this pair is that a rate card is an input to cost, not the cost. Effort settings and output token volume decide the invoice, and the only way to know which model is cheaper for your workload is to run your own tasks through both at pinned effort levels and count dollars per completed task. Building that evaluation harness and the routing layer behind it is the kind of applied AI work we do at Contra Collective.