Kimi K3.1 Pro vs Grok 4.7: Same Accuracy, a 72% Bigger Bill

On a private 43-question data-analytics exam built over a synthetic wind-farm warehouse, Kimi K3.1 Pro and Grok 4.7 both scored 42. Neither model had seen the questions, and the data was fabricated so no answer could have been memorised. Same score, same rank — tied for third of 34 models tested. The invoices were not the same: Kimi's run cost $13.47 and Grok's cost $7.85.

That pair of numbers is the most useful thing anyone has published about this matchup, because it is the only head-to-head where both models were run by the same people, on the same day, against the same answers. Almost everything else in circulation is either a vendor benchmark produced under a different scaffold or a comparison that quietly substitutes Kimi K3 for Kimi K3.1 Pro. So this piece starts with what the neutral tests show, then separates the differences that are real from the ones that are artefacts of who ran the test.

Two models, two very different bets

Kimi K3.1 Pro arrived on October 23, 2026 as a post-trained successor to Kimi K3 rather than a new architecture. Moonshot AI kept the 2.8-trillion-parameter mixture-of-experts backbone, the 1,048,576-token context window and native image and video input, and spent the release on reinforcement learning across agentic and professional domains. The full weights are promised by November 4, 2026.

Grok 4.7 shipped a month earlier, on September 21, 2026, from xAI. It is proprietary, sits on a reported ~2.1-trillion-parameter base that xAI has not confirmed, carries a 500,000-token input window with a 450,000-token output ceiling, and keeps the price of Grok 4.6. Its knowledge cutoff is June 2026, with supplemental training through August.

• Price — Kimi K3.1 Pro $2.00 input / $10.00 output per million tokens; Grok 4.7 $2.00 / $6.00, rising to $4.00 / $12.00 once a prompt passes 200K tokens.

• Cached input — $0.20 on Kimi against $0.50 on Grok, an inversion that matters more than the headline price for agents that re-read the same repository every turn.

• Context — 1,048,576 tokens against 500,000. Kimi also permits up to 943,718 output tokens in a single response, which is less a feature than a statement about what the model is built for.

• Modalities — text, image and video in, text out on Kimi; text, image and file in, text out on Grok.

• Reasoning control — Kimi reasons by default with selectable effort; Grok exposes low, medium, high and xhigh, and does not let you switch reasoning off.

• Openness — weights promised by November 4 on Kimi; Grok is API-only and will stay that way.

The neutral ground, and how thin it is

Plotly's exam is the cleanest comparison available. It is a private 43-question test over a synthetic wind-energy fleet warehouse, split into analytics questions, ambiguous requests, impossible requests, statistics traps, data-quality checks and five planted patterns a good analyst should find unprompted. Kimi K3.1 Pro passed 42, up from 35 for Kimi K3. Grok 4.7 also passed 42. Both trail GPT-6.1 Sol and GPT-6 Astra at 43.

The detail behind the tie is where the two models diverge. Kimi found all five planted patterns and caught all three data-quality problems — duplicate rows, a missing final day, and the difference between a decommissioned turbine and one genuinely producing zero. Its single miss was a boundary error: it counted 2,775 turbine-days from before a firmware installation as "other firmware", which moved an average by about 0.0003 and fell just outside the test's tolerance. That is a precision failure at the edge of a definition, not a reasoning failure.

A second neutral test points the other way. On DualEntry's accounting evaluation, Grok 4.7 finished tied for second at 88.1% while Kimi K3.1 Pro was fourth at 86.1%. Kimi was the cheapest model per task on that run at $0.079, but also one of the slowest at 125.6 seconds per task. Two neutral tests, two near-ties, and the ordering flips between them. That is what a genuine peer comparison looks like, and it is what most launch-day coverage does not.

Where the differences are real

Writing is not close. ToneBench, which scores drafts against ten briefs with four judges per draft, puts Kimi K3.1 Pro at 2,205 Elo and seventh of 83 configurations, against Grok 4.7 at 1,904 and twenty-fifth. The 95% intervals do not overlap: 2,166–2,246 against 1,855–1,962. Kimi won nine of ten scripts and seven of nine sub-metrics, with the widest gaps on length adherence (95.4 against 78.7) and visual cues (84.9 against 78.8). Grok's wins were narrow and confined to tone and YouTube conventions. At $0.251 against $0.215 per script, this is not a cost decision.

Coding and agentic work is where the evidence gets messy, and where most comparisons mislead. xAI reports Grok 4.7 at 37.6–38.0% on Terminal-Bench 4.0. Moonshot reports Kimi K3.1 Pro at 65.7% on the same benchmark. Those numbers sit 28 points apart and they do not mean Grok is half the model, because Terminal-Bench scores move by tens of points with the harness, the time limit and the attempts allowed. A 65.7% and a 38.0% produced in different labs are two different measurements wearing the same name.

The independent coding evidence that does exist favours Grok, but it is evidence about Kimi K3 rather than K3.1 Pro. On SWE-bench Verified run through a standard mini-agent scaffold, Grok 4.7 scored 95.6% against Kimi K3's 93.4%. Nobody has published the equivalent figure for K3.1 Pro. Moonshot's own long-horizon numbers — 61.2% on Kimi Code Bench 3.0 Long-horizon against 55.95% for K3, and 51.7% on SWE Marathon against 45.6% for Claude Opus 5.5 — suggest the post-training helped, but that is the vendor grading its own homework on a benchmark it built.

Long context is the one capability gap that needs no benchmark. A million tokens against five hundred thousand is a factor of two, and Kimi's cache-hit price is less than half Grok's. If the workload is a monorepo, a contract set or a year of support tickets, the comparison is decided before any score is read.

Why the invoice diverges

The 72% cost gap on the Plotly exam is not explained by the rate card alone. Kimi's output price is 67% higher, so on identical token counts it would cost more — but not 72% more on a test where the two models produced the same answer quality. The rest comes from how much they write and how many calls they make to get there.

Kimi averaged about 7.5 model calls per question and a median 2.5 minutes per question, the second-slowest of 34 models; roughly 90% of that time was model response rather than SQL execution. On hidden-pattern questions it made about 73 tool calls. Grok's independent testing elsewhere has found it consumes roughly twice the output tokens per task that Grok 4.6 did, which cuts the other way. The net effect on this exam was that Grok spent fewer dollars per correct answer despite burning more tokens per step.

The arithmetic for a production agent looks like this. Take a task that reads 100K input tokens, writes 8K of output, and runs 20 turns.

• Kimi K3.1 Pro, no cache — 20 turns at $0.20 input plus $0.08 output is about $5.60 per task.

• Grok 4.7, no cache — the same shape at $0.20 input plus $0.048 output is about $4.96 per task.

• Kimi K3.1 Pro, 70% cache hit — input falls to $0.074 a turn, taking the task to roughly $3.08.

• Grok 4.7, 70% cache hit — input falls to $0.095 a turn, taking the task to roughly $2.86.

So Grok wins the agent case even at a high cache rate, because output tokens dominate and it charges 40% less for them. Kimi only overtakes when output is small relative to input. Take a retrieval call with 100K input and 1K output at 70% cache: Kimi costs $0.084 and Grok $0.101, a 17% win for Kimi. The crossover is not a cache percentage — it is an input-to-output ratio. Read-heavy work goes to Kimi, write-heavy work goes to Grok, and the honest advice is to measure your own ratio rather than trust either vendor's example.

Speed changes the shape of the system, not just the bill

Kimi K3.1 Pro's median time to first token on our own measurements is 6.9 seconds at 48 tokens per second. Grok 4.7, in third-party testing, runs around 1.3 seconds to first token and roughly 61 tokens per second. For a chat product that difference is noticeable. For an agent making 20 sequential calls it is the difference between a task that feels instant and one that feels broken, and it is the real reason several coding-tool comparisons hand the win to Grok despite Kimi's stronger writing scores.

Reasoning control compounds it. Grok lets you dial effort from low to xhigh and charges accordingly; at its low setting ToneBench measured 1,815 Elo at $0.064 per script, a quarter of the medium setting's cost. Kimi reasons by default and its effort setting is a blunter instrument, with sampling fixed at temperature 1.0 and top_p 0.95 — values it rejects alternatives to. If a pipeline depends on low-temperature determinism, that is a compatibility question to settle before benchmarking anything.

Which one belongs in production

Choose Kimi K3.1 Pro when the work is long-context and document-heavy: contract review, financial research, multi-file repository analysis, and data-quality auditing where the model must notice what is missing rather than answer what is asked. The 1M window, the cheap cache and Moonshot's reported fall in hallucinated claims from 23.4% to 8.3% with tools all point the same way. It is also the only one of the two that will be downloadable, which matters if the workload cannot leave your infrastructure.

Choose Grok 4.7 when latency is part of the product: interactive coding assistants, IDE agents, customer-facing tool use where a six-second pause reads as a defect. It is cheaper on output, faster to first token, and the independent coding evidence — thin as it is — sits on its side. Its 500K window is not a limitation for most repository work, though it is for some.

The uncomfortable answer for anyone who wants a single winner is that the two models tie on the only neutral test both have taken, and the tiebreak is your input-to-output ratio and your latency budget. Those are properties you can measure in an afternoon, and no launch post can tell you what they are.

Running both without a second contract

This is the one matchup where trying both is cheap to arrange. OrcaRouter carries kimi/kimi-k3.1-pro and grok/grok-4.7 on the same key, at provider list price with no markup, so the $2.00, $10.00 and $6.00 above are what you are billed rather than a reseller's version of them. That matters more than usual here, because the decision depends on measured cache behaviour and latency inside your own application — which means you need both models behind one endpoint and one bill to compare them fairly.

It also changes the fallback story. Automatic failover across providers means a rate limit or an outage on one model does not take the feature down, and the routing DSL lets you send the long-context path to Kimi and the interactive path to Grok inside a single policy instead of maintaining two integrations. If you are still deciding, that is a way to run the evaluation on a slice of production traffic rather than picking from a benchmark table.

What would change this verdict

Two things, both dated. The first is November 4, when Moonshot has promised the full weights. An open-weight Kimi K3.1 Pro at this quality changes the calculus for anyone with a compliance reason to self-host, and it is also when independent labs can run the model without an API budget.

The second is a proper agentic head-to-head. Every coding comparison in circulation either uses Kimi K3 instead of Kimi K3.1 Pro, or compares two vendor harnesses as if they were one measurement. The moment someone runs both models through the same scaffold on the same tasks — the way Plotly did for data analytics — the coding question stops being contested. Until then, the defensible position is the boring one: Grok 4.7 for speed and interactive coding, Kimi K3.1 Pro for long context, writing and document work, and both of them measured on your own traffic before either gets a production path.