Grok 4.7 nearly doubles Terminal-Bench at the same $2/$6 price — and the number is already contested

What actually changed

• A bigger base model at the same price. xAI released Grok 4.7 on 21 September 2026 with a 2.1T-parameter base, up roughly 40% from Grok 4.6's 1.5T, and left list pricing at $2 per million input tokens and $6 per million output tokens.
• Four self-reported benchmark jumps, the largest of them on Terminal-Bench 4.0, where xAI's harness puts 4.7 at 38.0% against 4.6's 20.3%.
• That headline number is contested. Artificial Analysis, measuring independently, reports roughly 26% on the same benchmark.
• The price may not be the price. VentureBeat flags high token consumption, which can erode the advertised advantage once you count tokens actually burned on real work.
• It is not a clean sweep. Fable 5.1 Max still leads several coding and long-task evaluations.

The pattern of the past year has been that each new frontier release arrives with a price cut attached, and builders have learned to read the cut as the headline. Grok 4.7 breaks that pattern in a way that is worth noticing: xAI did not cut the price and did not raise it either. It shipped a substantially larger model and held the rate card still. For teams in Bengaluru and Bristol alike who have been rewriting their cost models every six weeks, an unchanged rate card is unusually restful news.

The restfulness lasts until you look at the benchmark table, which is where this release becomes a case study in why you should never migrate a production workload on the strength of a vendor's own chart.

The benchmark table, and how to read it

Every figure in the table below is xAI's own self-reported result, measured on xAI's own harness. That is not an accusation of bad faith; it is the ordinary condition of model launch material across every lab. But it matters enormously for agentic benchmarks, where the scaffold around the model contributes as much variance as the model itself.

Benchmark	Grok 4.7	Grok 4.6	Change
CursorBench 4.0	46.3%	40.4%	+5.9 points
DeepSWE v1.1 (high effort)	71.0%	65.2%	+5.8 points
EEBench	64.0%	53.0%	+11.0 points
Terminal-Bench 4.0	38.0%	20.3%	+17.7 points

Three of those four movements are the kind of generational step you would expect from a 40% larger base: meaningful, single-digit-to-low-double-digit gains that would show up as a modest but real improvement in a production agent. The fourth is an outlier. Terminal-Bench 4.0 nearly doubles, and a near-doubling on a hard agentic benchmark is the sort of claim that ought to attract scrutiny before it attracts migrations.

Where the two numbers diverge

It attracted exactly that. Artificial Analysis, which runs its own independent evaluation harness across frontier models, measures Grok 4.7 on Terminal-Bench 4.0 at approximately 26% — twelve points below xAI's 38.0%. Both organisations are reporting in good faith about the same benchmark and the same model, and both numbers are presumably reproducible inside their respective setups.

This is not a scandal. It is a structural property of agentic evaluation that builders need to internalise, because it will recur with every release from every lab for the foreseeable future. Terminal-Bench scores an agent's ability to complete tasks in a shell environment, and the result depends on a stack of choices sitting outside the model weights: which scaffold drives the loop, which tools are exposed and how their schemas are written, how many retries are permitted, how failures are classified, what reasoning-effort setting is used, how long the agent is allowed to run, and how partial completion is scored. Change any one of those and the number moves. Change several and it moves a lot.

The practical consequence is straightforward. If your agent runs on a scaffold closer to xAI's, you may see something nearer the optimistic figure. If it runs on a third-party harness with a different retry policy and a tighter tool surface, you may see something nearer the conservative one. Neither published number tells you which. Only your own run does. We have written before about benchmarking coding agents on your own repository rather than on SWE-bench, and this release is the strongest argument yet for that discipline.

The price that might not be the price

The second thing worth slowing down for is the cost story, which looks unambiguous and is not. At $2 per million input tokens and $6 per million output tokens, Grok 4.7 sits at an aggressive price point, and it inherits that point unchanged from a smaller predecessor. On paper, that is a straight improvement: more capability per pound, per rupee, per dollar.

VentureBeat's caveat is that Grok 4.7 exhibits high token consumption, and that this can erode the headline price advantage in real workloads. The mechanism is easy to state and easy to miss. A model that reasons at greater length, calls more tools, or takes more turns to converge will burn more tokens reaching the same finished outcome. Per token it is cheap. Per finished task it may not be.

Here is the arithmetic in a form that reads the same in Chennai and in Manchester. Suppose a workload consumes 50 MTok of input and 10 MTok of output in a month. At $2/$6 that is $100 plus $60, or $160. Now suppose the newer model completes the same volume of work but consumes 1.5 times the tokens along the way: 75 MTok in and 15 MTok out, which is $150 plus $90, or $240 — a 50% increase on an identical rate card. The multiplier in that example is illustrative, not measured; the point is only that the rate card and the bill are two different things, and that only your telemetry connects them.

What this does not change

Grok 4.7 is not a leaderboard sweep, and xAI has not claimed it is. Fable 5.1 Max continues to lead on several coding and long-task evaluations, so teams that have standardised on it for repository-scale work do not have an obvious reason to move on the basis of this release alone. What Grok 4.7 offers is a different trade: a large capability step at a price point that was already among the more aggressive, with no rate increase to absorb.

That trade is genuinely attractive for high-volume, cost-sensitive work — the kind of pipeline where a mid-tier model handles the bulk of traffic and a frontier model handles escalations. It is much less obviously attractive for the long-horizon coding tasks where the independent Terminal-Bench measurement suggests the gap between claim and delivery is widest.

On availability: the model is reachable through the Grok API, Grok Build, Cursor, a range of third-party coding harnesses and cloud platforms. That breadth is itself a reason to be careful. The same model behind five different harnesses is, for practical purposes, five different products with five different cost profiles.

The re-qualification drill for this week

The honest answer to "should we switch to Grok 4.7?" is that nobody outside your team can give you one, because the two published numbers for its flagship benchmark differ by twelve points and neither was measured on your stack. What you can do this week is run a short, disciplined re-qualification and produce a number that is actually about you. Five steps, and none of them need more than a day.

1. Freeze a task set from your own logs. Pull 40 to 60 real tasks that your current agent has already attempted in production, with their outcomes recorded. Real traffic, not curated examples. This set is your benchmark and it does not change between runs.
2. Pin the harness. Same scaffold, same tool definitions, same retry policy, same timeout, same effort setting for both models. If you change the harness and the model together, your result tells you nothing. This is the step that most teams skip, and it is the one the Terminal-Bench discrepancy exists to warn you about.
3. Record tokens, not just outcomes. Capture input tokens, output tokens, tool-call count and wall-clock time per run alongside pass/fail. Without the token columns you cannot test the VentureBeat caveat on your own data, and that caveat is the one most likely to bite.
4. Compute cost per successful task for both models. Total spend divided by passes. Publish the two numbers side by side to your team with the task set and the harness configuration attached, so the result is reproducible next month when the next release lands.
5. Route rather than replace. If Grok 4.7 wins on a subset — say, short-turn tool use — send that subset to it and leave the rest where it is. Wholesale swaps are how teams discover regressions in production. Shadow the new model against live traffic before it serves any of it.

If your team has not built this drill before, the mechanics are the same regardless of which vendor prompts the review: our walkthrough of re-qualifying a model after a price change lays out the three-day version, and shadow and canary deploys for model upgrades covers the rollout half. Teams whose coding agent picks its own default model should also read what happens when that default changes under you, because a model swap you did not initiate is still a model swap you own.

The takeaway

Grok 4.7 is a real step: a 40% larger base model, four improved benchmarks on xAI's harness, and a rate card that did not move. Those are facts and they are good ones. The headline Terminal-Bench figure, though, has an independent counter-measurement twelve points below it, and the price advantage carries a token-consumption asterisk that only your own telemetry can resolve.

None of that makes the release less interesting. It makes the release a reminder that the unit of decision for a working builder is not the leaderboard — it is cost per successful task, measured on your own workload, in your own harness, this week. Run the drill. Keep the number. You will need it again in six weeks.

Release details at x.ai/news/grok-4-7. Independent measurements at artificialanalysis.ai. Our earlier coverage of the previous release is at Grok 4.6 hits the frontier at a fraction of the price.