Grok 4.5

Grok 4.5 (API model ID grok-4.5) is xAI's frontier model, released 2026-07-08 as the first model jointly trained with Cursor: xAI used "trillions of tokens of Cursor data that capture real developer-agent interactions" during training, rather than static code corpora alone (Cursor blog). Press coverage of the launch consistently identifies the releasing entity as SpaceXAI — treated here as a rebrand or corporate restructuring of xAI rather than a separate company, since docs and press use the two names interchangeably for the same product line (TechCrunch, 2026-07-08). Elon Musk called it an "Opus-class model, but faster, more token-efficient and lower cost," and said internally it is "roughly comparable to Claude Opus 4.7, but much faster" — one generation behind Claude Opus 4.8, the model this page's benchmark tables actually compare it against (TechCrunch, 2026-07-08). Cursor positions it as broader than a coding specialist — "our most intelligent model, and the first... built for more than software engineering," targeting multistep problems across coding, data science, finance, and legal work (cursor.com/grok-4-5). It shipped at launch in Grok Build (xAI's terminal coding agent), on all Cursor plans, and via the SpaceXAI API console; it is not available in the EU at launch, with availability expected mid-July 2026 citing EU AI Act compliance work (cursor.com/grok-4-5).

Key specs

Values as of release (2026-07-08), re-verified 2026-07-12 against vendor docs and the models.dev and LiteLLM registries, which agree with each other on context window, output cap, and pricing. The 500K-token context window and sub-Opus pricing are the two numbers that dominate cost math for high-volume agent loops.

Spec	Value	Source
Context window	500,000 tokens	docs.x.ai/developers/models
Max output tokens	Not published on the vendor model page; the models.dev and LiteLLM snapshots (2026-07-12) both list 500,000, mirroring the context-window figure rather than a documented output cap — treat as unconfirmed by xAI	models.dev; LiteLLM
Knowledge cutoff	2026-02-01	docs.x.ai/developers/grok-4-5
Reasoning effort	low / medium / high, default high	docs.x.ai/developers/grok-4-5
Pricing, standard	$2 / MTok input, $6 / MTok output	docs.x.ai/developers/grok-4-5; confirmed by cursor.com/grok-4-5
Prompt caching	$0.50 / MTok cache read	models.dev; LiteLLM
Pricing, Cursor "Fast" variant	$4 / MTok input, $18 / MTok output	cursor.com/grok-4-5
Architecture	Mixture-of-experts (MoE); no published parameter count	cursor.com/blog/grok-4-5
Modalities	Text + image + PDF input; text output	models.dev

MarkTechPost reports a serving speed of roughly 80 tokens/sec and training on "tens of thousands of NVIDIA GB300 GPUs," both as of the 2026-07-08 launch and both single-source claims not corroborated in vendor docs (MarkTechPost, 2026-07-08). xAI's official docs publish no parameter count. Secondary coverage reports roughly 1.5T parameters as an xAI-stated ("self-reported") figure, but it appears in no vendor doc and has no independent audit, so it should not be treated as a confirmed spec (digitalapplied.com).

Benchmark results

Benchmark	Grok 4.5	Claude Fable 5	GPT-5.6 Sol	Claude Opus 4.8	GPT-5.5	Source
SWE-bench Verified (Vals AI harness)	86.6%	95.0%	96.2%	88.6%	82.6%	Vals AI leaderboard

As of 2026-07-12, from the Vals AI SWE-bench Verified leaderboard — a third-party harness, not xAI's own measurement. Grok 4.5 trails the other frontier models shown in this row except GPT-5.5 (82.6%), which it edges out.

The only other Grok 4.5 benchmark data found comes from xAI's own launch-day chart, relayed through secondary outlets because the original announcement page (x.ai/news/grok-4-5) returns HTTP 403 on every fetch attempt. None of the four rows below has an independent reproduction as of 2026-07-12, and the comparison figures quoted alongside Grok 4.5 in each row are themselves other vendors' self-reported numbers — apidog.com's own methodology note treats the whole set as vendor-vs-vendor self-report, not neutral benchmarking.

Benchmark	Grok 4.5	Comparison models (same table)	Source	As of
SWE-bench Pro, resolve rate	64.7%	Claude Fable 5 (max) 80.4%; Claude Opus 4.8 (max) 69.2%; GLM-5.2 62.1%; GPT-5.5 (xhigh) 58.6%	according to xAI, via apidog.com	2026-07-08
SWE-bench Pro, output tokens/task	15,954	Claude Opus 4.8 (max): 67,020 (~4.2x more)	according to xAI, via apidog.com	2026-07-08
Terminal-Bench 2.1, resolve rate	83.3%	Claude Fable 5 (max) 84.3%; GPT-5.5 (xhigh) 83.4%; Claude Opus 4.8 (max) 78.9%	according to xAI, via apidog.com, cross-confirmed by officechai.com	2026-07-08
SWE-bench Multilingual, resolve rate	78.0%	Claude Opus 4.8 (max) 84.4%; GPT-5.5 (xhigh) 77.8%	according to xAI, via officechai.com	2026-07-08
"SWE Marathon" (xAI-named eval), resolution rate	29.0%	Claude Opus 4.8 (max) 26.0%; Claude Fable 5 (max) 24.0%; Claude Opus 4.7 (max) 16.0%	according to xAI, via cryptobriefing.com	2026-07-08

Agentic behavior notes

• Cursor co-training. Trained on live developer-agent interaction data from Cursor rather than static code alone, per Cursor's own announcement (cursor.com/blog/grok-4-5).
• Self-disclosed benchmark contamination. Cursor disclosed that an earlier snapshot of its own codebase was accidentally included in Grok 4.5's training data, giving the model an unfair advantage specifically on CursorBench; Cursor excluded CursorBench from its own public comparisons as a result and says the contaminated data will be removed for future models. It is a first-party, self-disclosed integrity caveat — notable as a concrete instance of benchmark contamination rather than a theoretical concern (cursor.com/blog/grok-4-5).
• Cannot orchestrate sub-agent work. Independent testing found that when asked to decompose a task and delegate to child sub-agent processes, Grok 4.5 "lacked the nuance to split work intelligently and would hang when a sub-process stalled, never cleaning up." The same review calls it a strong single-agent tool for bug-finding, PR auditing, and function-writing, but "a generation behind" models built for orchestration, naming Claude Fable 5 and GPT-5.6 — while still calling it "an outstanding value default coding model" (BigGo Finance, relaying a third-party review).
• Grok Build is a separate capability from the model. Grok Build, xAI's terminal coding agent harness, runs on a dedicated coding model, grok-build-0.1 (256K-token context, 70.8% on SWE-bench Verified), not on Grok 4.5 itself (buildfastwithai.com). The harness supports up to eight parallel sub-agent processes across separate git worktrees with native MCP support. That parallelism is a harness feature of Grok Build, not evidence about Grok 4.5's own multi-agent reasoning — the orchestration weakness above is specifically about the model's own delegation logic (buildfastwithai.com).
• Caching and long loops. Vendor docs recommend setting a prompt_cache_key to route repeat conversations to the same server for reliable cache hits, and note that long agent loops on tool-heavy workloads benefit from context compaction (docs.x.ai/developers/grok-4-5).

Limitations

• No published system card as of 2026-07-12 (four days post-launch), a departure from xAI's own precedent with Grok 4 and Grok 4.1, both of which shipped model cards. Flagged as an open compliance gap by an AI-governance trade outlet — relevant for anyone evaluating Grok 4.5 for regulated or safety-sensitive agentic deployments (AI Governance Institute, 2026-07).
• Not available in the EU at launch, cited as EU AI Act-related work in progress; availability expected mid-July 2026 (cursor.com/grok-4-5; MarkTechPost).
• Weak sub-agent orchestration, per independent testing described above — the most consequential limitation for team-of-agents use cases: profile it as a single-agent worker model, not a lead/orchestrator model (BigGo Finance).
• Mixed, self-reported benchmark picture, and one contaminated benchmark. Outside the independently-sourced SWE-bench Verified row, every benchmark figure on this page traces back to xAI's own launch chart with zero independent reproduction as of 2026-07-12: Grok 4.5 trails Claude Fable 5 on SWE-bench Pro and Terminal-Bench 2.1, splits against Claude Opus 4.8 and GPT-5.5 on SWE-bench Multilingual, and leads an unverified "SWE Marathon" eval (see Benchmark results for the caveats on that row). Separately, CursorBench — one of the benchmarks in xAI's original comparison set — is vendor-disclosed as contaminated (apidog.com; cursor.com/blog/grok-4-5).
• Max output limit and parameter count are undocumented by the vendor (see Key specs) — plan capacity conservatively until xAI publishes either figure directly.

Related

• Claude Opus 4.8 — the generation Musk pitched Grok 4.5 against as "Opus-class," and the model Grok 4.5 trails on SWE-bench Verified; Musk's specific "roughly comparable to" quote actually names Claude Opus 4.7 (see lead paragraph).
• Claude Fable 5 — ahead of Grok 4.5 on SWE-bench Verified, SWE-bench Pro, and Terminal-Bench 2.1 in the sourcing for this page; not measured against it on SWE-bench Multilingual or "SWE Marathon."
• GPT-5.5 — co-occurs in most benchmark comparisons found for this release (all but the xAI-named "SWE Marathon" row, which omits it); Grok 4.5 beats it on SWE-bench Multilingual and roughly matches it on Terminal-Bench 2.1.
• SWE-bench — the only benchmark family with an independently-sourced score for Grok 4.5 in this cluster; the SWE-bench Pro and Multilingual figures above are xAI self-reports only.