GROK 4.5 — CONFIRMED SPECS AND PRICING
● Launched: July 8 (developers via Grok Build, Cursor, API) / July 9 (public on grok.com and X app)
● API string: grok-4.5
● Pricing: $2/$6 per million tokens · cached input: $0.50/M (75% discount) · above 200K tokens: higher-context surcharge applies
● Context window: 500K tokens — step DOWN from Grok 4.3's 1M. Critical limitation for large codebase work.
● Speed: ~80 tokens/second
● Architecture: 1.5T parameter V9 MoE (mixture-of-experts) — not disclosed officially but Musk-confirmed
● Training: Cursor developer session data (debugging traces, multi-file diffs, user corrections) + RL on engineering tasks
● Artificial Analysis Intelligence Index: #4 of 168 models (score 54) — above all Gemini models and all open-weight models
● EU: Not available at launch. Expected mid-July 2026.
The Complete Benchmark Picture
| Benchmark |
Grok 4.5 |
Claude Opus 4.8 |
Fable 5 |
GPT-5.5 |
Sonnet 5 |
| SWE-bench Pro |
64.7% |
69.2% ✓ |
80.4% |
58.6% |
63.2% |
| DeepSWE 1.0 (provider harness) |
62.0% ✓ vs Opus |
55.75% |
66.1% |
64.3% |
— |
| DeepSWE 1.1 (neutral harness) |
53% |
59% ✓ |
70% |
67% |
— |
| Terminal-Bench 2.1 |
83.3% |
78.9% |
84.3% |
83.4% |
80.4% |
| SWE Marathon (pass@1) |
29.0% ✓ |
26.0% |
24.0% |
— |
— |
| Artificial Analysis Index |
54 (#4) |
56 (#3) |
64.9 (#1) |
— |
— |
| Output tokens per SWE task |
15,954 (4.2x fewer) |
67,020 |
— |
— |
— |
Note: DeepSWE 1.0 uses each provider's own harness — Grok 4.5 wins here partly due to home-field advantage. DeepSWE 1.1 uses a neutral harness — Grok 4.5 trails Opus 4.8 by 6 points. The neutral harness result is more reliable for cross-model comparison. Also: Cursor acknowledged an earlier snapshot of its codebase was accidentally included in Grok 4.5's training data, potentially inflating CursorBench results specifically.
The Token Efficiency Story — Why It Changes the Cost Math
The most commercially significant Grok 4.5 number is not a benchmark score — it is the token efficiency figure. Grok 4.5 resolves tasks using an average of 15,954 output tokens against 67,020 for Opus 4.8 on SWE-Bench Pro — a 4.2x gap. This directly affects cost per completed task, not cost per token. Grok 4.5 is priced at $6/M output versus Opus 4.8 at $25/M output — a 4.2x price gap. But if Grok 4.5 uses 4.2x fewer output tokens per task, the effective cost per completed task may be roughly equivalent even though Grok 4.5 has lower absolute benchmark scores. The token efficiency advantage only matters if Grok 4.5's outputs are correct — token-efficient wrong answers cost less but produce worse results. Validate on your specific workloads before switching production traffic.
The 500K Context Window — The Real Limitation
Grok 4.5 has a 500,000-token context window. That is smaller than Grok 4.3's 1 million-token context listing, so do not assume context grew with the new model. 500K is approximately 375 pages — sufficient for most individual coding files but insufficient for large-codebase analysis or long-document synthesis tasks. For teams that relied on Grok 4.3's 1M context, this is a meaningful step backward. If your workload requires feeding entire repositories or large document sets into a single context, Grok 4.5 is the wrong upgrade path from Grok 4.3. Claude Sonnet 5 (1M context) or GPT-5.6 Sol/Terra (1.05M context) are better options for those workloads.
The Honest Verdict — What Musk's "Opus-Class" Claim Actually Means
Musk's comparison to Anthropic's Opus line was precise in a follow-up: "Grok 4.5 is roughly comparable to Opus 4.7, but much faster." That last detail matters. Musk is benchmarking against Opus 4.7, not 4.8, which the independent data broadly supports. Artificial Analysis ranks Grok 4.5 at 54 on its Intelligence Index versus Opus 4.8 at 56 — close but not ahead. On SWE-bench Pro (the gold standard for agentic coding), Opus 4.8 leads Grok 4.5 by 4.5 points (69.2% vs 64.7%). Grok 4.5 leads on SWE Marathon (29.0% vs 26.0%) and #1 on agentic tool-use.
Choose Grok 4.5 if: High-volume agentic coding, cost-per-task is your primary metric, token efficiency matters more than absolute benchmark score, workloads fit within 500K context. At $2/$6/M it is the best intelligence-per-dollar coding model available in July 2026.
Choose Claude Opus 4.8 instead if: Absolute coding quality is the priority (leads on SWE-bench Pro, DeepSWE 1.1), 200K context is sufficient, you trust independently replicated benchmarks over provider-harness results.
Choose Claude Fable 5 if: You need the highest coding ceiling (80.4% SWE-bench Pro, 84.3% Terminal-Bench), can afford $10/$50/M, and the 200K context window fits your workloads. Available on Max, Team, Enterprise with credits.
EU teams: Grok 4.5 is not available in the EU at launch. Expected mid-July 2026. Use Opus 4.8, Fable 5, or GPT-5.6 in the meantime.
Sources: Axios, Fello AI, Roo.beehiiv, eesel.ai, Kingy.ai, MLQ News, Awesome Agents, technology.org, codersera.com — July 8-9, 2026 · Related: GPT-5.6 Sol/Terra/Luna full review → · Model comparison hub → · SuperGrok vs Claude Pro vs ChatGPT Plus →