Grok 4.7: What xAI's CursorBench and DeepSWE Gains Actually Mean for Coding Teams
Grok 4.7 landed September 21, 2026 at the same two dollar input, six dollar output pricing as Grok 4.6, which is the detail that makes the release worth a second look. CursorBench 4.0 moves from 40.4 to 46.3 percent, DeepSWE v1.1 moves from 65.2 to 71.0 percent, and the model is already rolling out inside Cursor and GitHub Copilot. xAI credits a larger base model and a longer reinforcement learning run weighted toward difficult, hours long tasks. Here is what that training change actually bought, and what it did not.
Grok 4.7: What xAI's CursorBench and DeepSWE Gains Actually Mean for Coding Teams
xAI shipped Grok 4.7 on September 21, 2026, and the first thing worth noticing is not a benchmark number. It is the price. Grok 4.7 holds the exact same $2 per million input tokens and $6 per million output tokens that Grok 4.6 shipped at, through the xAI API, Grok Build, Cursor, and now GitHub Copilot, which began rolling the model out to Copilot users the same week. A frontier lab shipping a capability upgrade at unchanged pricing is the more interesting story than any single benchmark line, because it means the gains below are not funded by a price increase you will eventually pay for indirectly.
xAI attributes the release to three training changes: a larger base model, a longer reinforcement learning run, and a shift in what that RL run weights, specifically more emphasis on difficult tasks that take hours to complete rather than short, easily rewarded ones. The company also says self verification and long context management both improved as a result. Those are training process claims, not benchmark claims, and the benchmark deltas below are the closest thing to an independent check on whether they translated into anything measurable.
What actually moved
| Benchmark | Grok 4.6 | Grok 4.7 | Delta |
|---|---|---|---|
| CursorBench 4.0 (software engineering) | 40.4% | 46.3% | +5.9 points |
| DeepSWE v1.1 | 65.2% | 71.0% | +5.8 points |
| AA Briefcase (office work, Elo) | Not directly reported | 1,657 | Not directly comparable |
| Pricing (input / output per 1M tokens) | $2 / $6 | $2 / $6 | Unchanged |
Both coding benchmarks move by roughly the same margin, just under six points, which is a meaningfully larger jump than the single digit or sub one point gains that have characterized most point release upgrades across labs this year. A near six point gain on a software engineering benchmark from one release to the next is the kind of number that would normally accompany a full version jump rather than a 4.6 to 4.7 increment, and it is consistent with xAI's framing that this release leaned harder on reinforcement learning against difficult, long horizon tasks specifically, the category CursorBench and DeepSWE are built to stress.
Reading the gain honestly, not as a leaderboard headline
CursorBench 4.0 and DeepSWE v1.1 are not the same benchmarks we have used elsewhere on this site to evaluate Claude Code, Codex, and Fable 5.1, which mostly report against Terminal-Bench 4.0 and SWE-bench Pro. That is worth being explicit about, because the honest answer to "how does Grok 4.7 compare to Claude Fable 5.1 or GPT-6 Astra on coding" is that xAI has not published a head to head number on a shared benchmark yet, and neither have independent evaluators as of this week. What you actually have is a same lab, same benchmark, release over release comparison, which tells you the direction and magnitude of xAI's own progress but not where Grok 4.7 lands relative to the rest of the field on the benchmarks that field actually reports.
That distinction matters more than it sounds. A vendor reporting its own gain on its own chosen benchmark is a real data point, but it is not the same evidentiary weight as a third party leaderboard where every model runs the identical task set under identical conditions, the kind of format we have flagged as the more trustworthy read in our coverage of Terminal-Bench 4.0's harness and model pairings. Until Grok 4.7 shows up on a shared leaderboard against Fable 5.1, GPT-6 Astra, and Opus 5 under one methodology, treat the CursorBench and DeepSWE numbers as evidence that Grok 4.7 is a real step up from Grok 4.6, not as evidence of where it ranks against competitors.
The AA Briefcase number is a different kind of claim
The 1,657 Elo score on AA Briefcase, a benchmark aimed at general office and knowledge work rather than pure coding, is reported without a direct Grok 4.6 comparison point in the material xAI published alongside the release. That makes it useful context for teams evaluating Grok 4.7 for broader agentic work beyond a coding harness, spreadsheet manipulation, document synthesis, multi step research tasks, but it is not evidence of improvement the way the CursorBench and DeepSWE deltas are, since there is no prior number to measure it against. Treat it as a snapshot of current capability on that task category, not a trend line.
What unchanged pricing actually signals
Holding pricing flat while shipping a real capability gain is a specific competitive move, and it lands differently depending on which seat you sit in. For a team already paying for Grok 4.6 access through Cursor or the xAI API, the upgrade path is essentially free: same cost structure, measurably better performance on the two coding benchmarks xAI chose to highlight. For a team evaluating which frontier model to standardize on, unchanged pricing from Grok 4.7 sitting alongside the $10 input and $50 output pricing on GPT-6 Astra, roughly five times Grok's rate, keeps xAI positioned as the value option in the frontier tier rather than trying to compete on capability alone. That positioning has been consistent through the Grok 4.x line, and 4.7 does not change it, it just raises the capability bar at the same price point.
The GitHub Copilot rollout is the detail worth tracking operationally. Copilot users gaining Grok 4.7 as a model option means the model reaches a very large population of working developers who were never going to set up a direct xAI API integration, and Copilot's model picker becomes a meaningful additional distribution channel for a benchmark gain that otherwise would have stayed inside teams already committed to the xAI ecosystem.
FAQ
Did Grok 4.7's price change from Grok 4.6? No. Both are priced at $2 per million input tokens and $6 per million output tokens. The capability gain shipped at the same cost.
How does Grok 4.7 compare to Claude Fable 5.1 or GPT-6 Astra on coding? xAI has not published a head to head result on a shared benchmark with those models yet. The CursorBench 4.0 and DeepSWE v1.1 numbers are release over release comparisons against Grok 4.6, not cross vendor comparisons.
What changed in training between Grok 4.6 and Grok 4.7? xAI reports a larger base model, a longer reinforcement learning run, and a shift toward weighting difficult, hours long tasks more heavily during that RL run, alongside improved self verification and long context management.
Where can I use Grok 4.7 today? Through the xAI API, Grok Build, inside Cursor, and as a rolling out model option in GitHub Copilot.
When this applies to your stack
If you are already running Grok 4.6 through Cursor or the xAI API, upgrading to 4.7 costs nothing extra and the benchmark deltas suggest a real gain on difficult, long horizon coding tasks specifically. If you are choosing a frontier model from scratch, wait for a shared leaderboard result before treating Grok 4.7's own benchmarks as a ranking claim against Fable 5.1 or GPT-6 Astra. Building the evaluation harness that tests a candidate model against your actual task mix, rather than relying on any single vendor's chosen benchmark, is the kind of agent engineering work we do at Contra Collective.
More from the lab.
vLLM vs SGLang: Choosing a Production LLM Inference Engine in Late 2026
vLLM and SGLang are both open source, both built to serve LLMs at production scale, and both frequently recommended for the exact same job, which is exactly why teams keep picking the wrong one. The real difference is not a benchmark number. It is which attention and caching architecture matches your actual traffic pattern, and most teams never check that before they commit infrastructure budget to one or the other.
Ollama's MLX Backend Beats Metal by 93 Percent: Apple Silicon Local Inference Tested (October 2026)
Ollama's 0.19 release notes say it plainly: on Apple Silicon, Ollama is now built on top of Apple's MLX framework to take advantage of its unified memory architecture, rather than the Metal backed llama.cpp path it shipped for years. On an M4 Max running Qwen3.5, that swap took decode speed from 57.8 tokens per second to 111.4, a 93 percent improvement, and prefill from 1,154 to 1,810 tokens per second. The catch is in the fine print: the MLX runner is in preview, it needs 32GB or more of unified memory to activate at all, and as of this release it only accelerates one model family. Here is what actually changed, what it costs to turn on, and where it still falls back silently to the slower path.
DGX Spark vs Mac Studio M5 Ultra: Local LLM Inference Benchmarked (October 2026)
The M5 Ultra Mac Studio shipping in September 2026 is a quad die chip with 1.2 terabytes per second of memory bandwidth and up to 256GB of unified memory, with a 512GB configuration arriving in late October. Against a single NVIDIA DGX Spark, it is not close: on a 27B dense model, the Mac decodes at roughly four times the speed. Cluster two DGX Spark units together over NVIDIA's GX10 networking instead, and the comparison flips for a different workload entirely, concurrent multi user serving. Neither machine is the right answer for every workload, and the decision comes down to whether you are serving one user or several at once.