SpaceXAI releases Grok 4.6: same price as 4.5, mixed results on its own benchmarks

SpaceXAI released Grok 4.6 on 12 August, a follow-up to Grok 4.5 that the company positions around long-running agents — work that has to hold together across many steps, whether that is moving through a codebase, researching an unfamiliar topic, or taking a product idea to a first working version — and around interactive and visual output. The list price is unchanged from Grok 4.5: $2 per million input tokens and $6 per million output tokens.

Every performance number published with the release is SpaceXAI's own. That is worth holding onto, because the benchmark table in the announcement is the most informative part of it, and it does not read entirely in the company's favour.

What the release claims

The headline claim is parity on the Artificial Analysis Intelligence Index, which SpaceXAI describes as a composite of nine benchmarks. Grok 4.6 scores 61 there, tying GPT-5.6 Sol Max at 61 and sitting one point behind Fable 5 Max at 62. Grok 4.5 High scored 56, so the jump over its own predecessor is five points. Read plainly, that is a claim of matching the strongest OpenAI model in SpaceXAI's own comparison while still trailing the Anthropic one — not a claim of leading the field. The table carries four columns, two of them SpaceXAI's own models; it is a selected comparison, not a leaderboard.

On the evaluations where SpaceXAI's table puts Grok 4.6 ahead, the margins are mostly narrow:

• GDPVal-AA v2: 1753, against 1741 for Fable 5 Max, 1728 for GPT-5.6 Sol Max and 1526 for Grok 4.5 High.
• AA-Briefcase: 1577, against 1574 for Fable 5 Max and 1502 for GPT-5.6 Sol Max.
• Harvey LAB (Vals): 15.8%, against 11.3% for Fable 5 Max and 2.5% for GPT-5.6 Sol Max.
• CursorBench v3.2: 69.9%, ahead of GPT-5.6 Sol Max at 67.2% but behind Fable 5 Max at 70.5%.

Where its own numbers put it behind

The same table shows Grok 4.6 losing on the two evaluations that look most like the long-running agent work the model is being sold for. On DeepSWE v1.1 it scores 65.9%, against 73% for GPT-5.6 Sol Max and 70% for Fable 5 Max. On Terminal-Bench v3.0 it scores 26%, against 34.6% and 34.1%. It also trails Fable 5 Max on FrontierCode v1.1 Extended (61.3% to 63.6%), APEX-Agents (57.5% to 59.2%) and APEX-SWE (56.4% to 58.8%).

Publishing a table you lose parts of is better practice than the alternative, and SpaceXAI deserves the credit for it. But it means the parity claim is an index-level result rather than something that survives evaluation by evaluation. A model pitched at agents that keep working across many steps is, by its own measurement, roughly eight points behind on the terminal benchmark.

There is a second caveat in the fine print. SpaceXAI says competitor figures are drawn from the other developers' published system cards or leaderboards, and are "the best of self-reported or publicly available results". That is not a controlled head-to-head: the comparison points were produced on different harnesses, at different reasoning-effort settings, on different dates. Cross-vendor benchmark tables assembled this way indicate a rough tier, not a ranking you should trust to a decimal.

The parts that show up on your API bill

SpaceXAI's public model documentation carries the details that matter more than the index score if you are actually running this thing. Grok 4.6 has a 500,000-token context window, but the docs set a long-context threshold at 200,000 tokens, and past that boundary the rates double: input goes from $2 to $4 per million, output from $6 to $12. The window you can use without a price change is 200K, not 500K. For agent loops that accumulate transcript, tool output and file contents, that threshold is the number to instrument against.

Two other details are easy to miss:

• Cached input is listed at $0.50 per million ($1.00 in the long-context band). The same docs list Grok 4.5's cached input at $0.30 per million, so cache hits got more expensive even though the headline rates held flat.
• The default reasoning effort is high. Supported settings are low, medium, high and xhigh; on a reasoning model, effort maps to billed output tokens, so the default is the expensive one. Grok 4.3, by comparison, defaults to low and supports turning reasoning off entirely.
• The documented batch discount for Grok 4.6 is 0%, against 20% for Grok 4.3.
• Function calling and structured outputs are both supported, and the announcement notes a fast variant at twice the price.

Put together, the flat headline price is real but partial. Same $2/$6 as Grok 4.5, more expensive caching, no batch discount, a cost cliff at 200K tokens, and a default that spends more reasoning tokens than you may intend.

Availability

Grok 4.6 is available from today in Cursor and in SpaceXAI's own Grok Build, through the API, and via OpenRouter, Vercel and Cloudflare. SpaceXAI is offering double the included usage inside Grok Build and Cursor for the first week — a trial subsidy that expires, and one to keep out of any cost model you build this week.

On training, the announcement says Grok 4.6 had a longer supplemental training run than Grok 4.5, using curated model-generated data and an improved optimizer and recipe, that Grok 4.5 was used to regenerate the supervised fine-tuning trajectories, and that reinforcement learning covered agentic tasks including kernel optimization, web development and computer-aided design. The safety section describes the company's widest pre-deployment testing suite to date plus third-party testing, without naming the testers or publishing results.

What we could not verify

We could not open any independent evaluation of Grok 4.6. Artificial Analysis — the evaluator whose index the parity claim rests on — along with OpenRouter, Cursor's own post and several outlets that covered the launch, were all unreachable from our environment, either blocked at the network layer or behind bot challenges we will not work around. This article therefore rests on SpaceXAI's announcement and its public model documentation, both company sources.

Treat 61 on the Intelligence Index as SpaceXAI's claim rather than an established result until a third party publishes its own run. The pricing, context window, threshold behaviour and feature flags are on firmer ground: those come from the API documentation the company has to honour in billing.

If you are choosing a model for an agent loop, the concrete deltas here are the same headline price as Grok 4.5, a 200K cost cliff inside a 500K window, a high default reasoning effort, no batch discount, and a vendor table that puts the model behind on the two most terminal-shaped coding evaluations. That is enough to justify running it against your own tasks. It is not enough to justify swapping on an index number.