Grok 4.5
Grok 4.5 is xAI's 1.5-trillion-parameter V9 MoE model, publicly launched July 8 at $2/M input - cheap, fast, and token-efficient, though neutral harness benchmarks put it well behind Fable 5 and Opus 4.8 on coding.
Updated
8 min read
Grok 4.5 launched publicly on July 8, 2026 - completing the trip from Elon Musk's June 28 X post ("early evals close to, perhaps exceeding Opus") to a real API you can actually call. Built on the V9 foundation architecture with 1.5 trillion parameters, it's the first V9 model at full scale in general availability, running as a mixture-of-experts rather than a dense transformer.
TL;DR
- 1.5T MoE model, 500K context (cut from Grok 4.3's 1M), $2/M input / $6/M output
- Benchmark story is split: provider-harness DeepSWE 1.0 puts it ahead of Opus 4.8; neutral-harness DeepSWE 1.1 puts it well behind both Opus 4.8 and Fable 5
- Token efficiency is the clearest edge - ~15,954 output tokens per SWE-Bench Pro task vs 67,020 for Opus 4.8
The private beta at SpaceX and Tesla ran from June 28 through July 7, with real engineering workloads - rocket path calculations, manufacturing software - feeding pass/fail RL signal back into training. Whether that live RL environment produced a model that matches the private-beta hype is what the public benchmarks now let us answer. The answer is: it depends heavily on who runs the evaluation.
One caveat Cursor disclosed at launch: an earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5's training data. This inflates scores on any CursorBench-derived evaluation. Third-party benchmarks like DeepSWE and SWE-Bench Pro use their own independent infrastructure and aren't affected.
Key Specifications
| Specification | Details |
|---|---|
| Provider | xAI (SpaceX AI division) |
| Model Family | Grok / V9 |
| Architecture | Mixture-of-experts (MoE) |
| Parameters | 1.5 trillion |
| Context Window | 500K tokens |
| High-Context Threshold | Higher rates above 200K tokens |
| Input Price | $2.00/M tokens |
| Cached Input | $0.50/M tokens |
| Output Price | $6.00/M tokens |
| Fast Variant (Cursor) | $4.00/M input / $18.00/M output |
| Release Date | July 8, 2026 |
| API Model ID | grok-4.5 |
| Speed | 80 TPS (xAI stated) / 91.3 TPS (Artificial Analysis) |
| Open Source | No |
| License | Proprietary |
The context window is worth flagging. Grok 4.5 ships with 500K tokens - half the 1M-token window on Grok 4.3. xAI hasn't publicly explained the reduction. For teams using Grok for document analysis or ingesting large codebases, this is a real constraint Grok 4.3 didn't have.
Benchmark Performance
The picture that emerges from public benchmarks is more nuanced than the pre-launch framing suggested.
| Benchmark | Grok 4.5 | Fable 5 | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| DeepSWE 1.0 (provider harness) | 62.0% | 66.1% | 55.75% | 64.31% |
| DeepSWE 1.1 (neutral harness) | 53% | 70% | 59% | 67% |
| Terminal Bench 2.1 | 83.3% | 84.3% | 78.9% | 83.4% |
| SWE-Bench Pro | 64.7% | 80.4% | 69.2% | 58.6% |
| Artificial Analysis Index | 54 (#4) | - | - | - |
| GDPval-AA v2 Elo | 1543 | - | - | - |
The gap between DeepSWE 1.0 and 1.1 tells the most important part of the story. On the provider-controlled harness, where xAI designs the scaffolding, Grok 4.5 beats Opus 4.8 by 6 points. On the neutral mini-swe-agent harness, Grok 4.5 drops 9 points and falls behind both Opus 4.8 and GPT-5.5. That swing is too large to chalk up to randomness; it suggests the model is sensitive to how the task scaffold is set up.
Terminal Bench 2.1 is the cleaner story - all four models cluster within 5.4 points, with Grok 4.5 at 83.3% nearly matching GPT-5.5 at 83.4%. SWE-Bench Pro ranks Grok 4.5 third out of four: behind Fable 5 (80.4%) and Opus 4.8 (69.2%) but ahead of GPT-5.5 (58.6%).
Token efficiency is where Grok 4.5 actually delivers on the private-beta promise. xAI reports an average of 15,954 output tokens per SWE-Bench Pro task - against 67,020 for Opus 4.8. At the per-token prices involved, that 4.2x efficiency gap changes the cost math significantly for high-volume agentic workloads. GPQA Diamond and MMLU-Pro scores haven't been published yet, so direct reasoning comparisons remain incomplete.
Key Capabilities
MoE Architecture and V9 Scale
Grok 4.5 is a mixture-of-experts model. At 1.5T total parameters, the active-parameter count per inference is substantially lower than a 1.5T dense model would be - which explains how xAI hits 80-91 TPS throughput at this scale. The same architectural pattern is used by Mistral's MoE models and DeepSeek.
The V9 foundation completed its primary training run on May 26, 2026. Grok 4.5 added Cursor session data as supplemental training: real developer workflows, debugger interactions, multi-file diffs, and user corrections from Cursor's user base. The dataset is intentionally broader than Cursor's own previous Composer 2.5 model. One caveat stands out: Cursor's launch post acknowledged that an earlier Cursor codebase snapshot was accidentally included in training, giving Grok 4.5 a dataset advantage specifically on CursorBench that doesn't carry over to neutral evaluations.
Grok Build Integration
During private beta, xAI ran the model through the Grok Build harness - autonomous code execution at SpaceX and Tesla, with pass/fail results feeding back into ongoing RL. That harness is now a public product: the Code API exposes grok-build-0.1 for autonomous multi-step coding tasks distinct from the main grok-4.5 chat model.
The token efficiency numbers flow directly from how Grok Build structures tasks. Rather than generating long verbose responses, the harness breaks work into many short tool calls - which hits the ~15K average output tokens per SWE-Bench Pro task. That's efficient when the short calls are accurate; it can build up errors across a long task chain when they aren't.
Grok 4.5's supplemental training drew on Cursor session data - real developer debugging workflows rather than static code repositories.
Source: unsplash.com
Regional and Platform Availability
At launch, Grok 4.5 is available on the xAI API, in Cursor (desktop, web, iOS, CLI, and SDK), through Microsoft Office add-ins, and via model gateways including OpenRouter. EU access wasn't included at launch - xAI cited regulatory compliance review and expects access around mid-July 2026. No specific date has been given.
Pricing and Availability
At $2/M input and $6/M output, Grok 4.5 is priced well below its closest performance competitors. Opus 4.8 costs $5/M input and $25/M output. Even accounting for Grok 4.5's lower benchmark scores on neutral evals, the 4.2x token efficiency advantage over Opus 4.8 means the effective cost per unit of work done is lower than the list price gap alone suggests.
| Model | Input $/M | Output $/M | Context | Artificial Analysis Index |
|---|---|---|---|---|
| Grok 4.5 | $2.00 | $6.00 | 500K | 54 (#4) |
| Grok 4.3 | $1.25 | $2.50 | 1M | ~53 |
| Claude Opus 4.8 | $5.00 | $25.00 | 1M | - |
| GPT-5.5 | Not confirmed | Not confirmed | Not confirmed | - |
| Claude Fable 5 | Not confirmed | Not confirmed | Not confirmed | - |
The comparison with Grok 4.3 is worth noting: at 53 vs 54 on the Artificial Analysis Index, the performance delta is small - yet Grok 4.5 costs 60% more per input token and offers half the context window. For teams whose workloads need long context rather than high-volume agentic throughput, Grok 4.3 is still a defensible choice within the xAI lineup.
The Cursor-specific fast variant at $4/M input and $18/M output is available for latency-sensitive workflows, with roughly double the speed at three times the output cost.
Cursor, now under SpaceX acquisition pending close in Q3 2026, built Grok 4.5 alongside xAI and ships it across desktop, web, iOS, and CLI.
Source: unsplash.com
Strengths
- Price-to-performance ratio - $2/M input is well below Opus 4.8 and most Fable 5 tiers
- Token efficiency: ~4.2x fewer output tokens than Opus 4.8 per agentic task on SWE-Bench Pro
- Fast inference: 80-91 TPS, viable for latency-sensitive pipelines
- Broad availability at launch: xAI API, Cursor, Microsoft Office add-ins, OpenRouter
- MoE architecture scales efficiently for high-volume deployments
Weaknesses
- Neutral harness benchmarks (DeepSWE 1.1, 53%) tell a different story than provider-controlled ones (DeepSWE 1.0, 62%)
- Context window cut from Grok 4.3's 1M to 500K with no public explanation
- CursorBench scores contaminated by accidental training data overlap
- EU unavailable at launch; no precise date for mid-July access
- GPQA Diamond and MMLU-Pro scores not yet published
- SWE-Bench Pro score (64.7%) trails both Fable 5 (80.4%) and Opus 4.8 (69.2%)
Related Coverage
- Predecessor: Grok 4.3 | Grok 4.3 Review
- Earlier models: Grok 4.20 | Grok 4
- Competitor: Claude Opus 4.8 - the benchmark target Grok 4.5 was measured against in private beta
- Leaderboards: Coding Benchmarks | SWE-Bench Leaderboard | Reasoning Benchmarks
FAQ
What is Grok 4.5?
Grok 4.5 is xAI's 1.5-trillion-parameter MoE model built on the V9 foundation, publicly launched July 8, 2026. Available on the xAI API, Cursor, and Microsoft Office add-ins at $2/M input tokens.
How does Grok 4.5 compare to Claude Opus 4.8 on coding?
On provider-controlled DeepSWE 1.0, Grok 4.5 (62%) beats Opus 4.8 (55.75%). On neutral DeepSWE 1.1, Opus 4.8 (59%) beats Grok 4.5 (53%). Grok 4.5 uses 4.2x fewer output tokens per task but trails Opus 4.8 on SWE-Bench Pro (64.7% vs 69.2%).
Why is the context window smaller than Grok 4.3?
Grok 4.5 ships with 500K tokens vs Grok 4.3's 1M. xAI hasn't explained the reduction publicly. Requests above 200K tokens also incur higher-context rates, so long-context workloads cost more per token than the headline list price suggests.
What is the CursorBench contamination issue?
Cursor disclosed that an earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5's training data. This gives Grok 4.5 a dataset advantage on CursorBench specifically. Third-party benchmarks - DeepSWE, SWE-Bench Pro, Terminal Bench - use their own infrastructure and aren't affected.
Is Grok 4.5 available in Europe?
Not at launch. xAI expects EU access around mid-July 2026, pending regulatory compliance review. No exact date has been published.
What is the difference between grok-4.5 and grok-build-0.1?
grok-4.5 is the standard chat and code model. grok-build-0.1 is the specialized Code API model for the Grok Build autonomous coding harness - it's designed for multi-step task orchestration with autonomous tool execution, not conversational use.
Sources:
- Introducing Grok 4.5 - xAI
- Introducing Grok 4.5 - Cursor Blog
- Grok 4.5 model page - OpenRouter
- xAI Developer Models Documentation
- SpaceXAI Releases Grok 4.5 - MarkTechPost
- Grok 4.5 benchmarks, pricing, context window - Kingy.ai
- SpaceX's Grok 4.5 launches at half the price of rivals - VentureBeat
- Grok 4.5 vs Claude Opus 4.8 - APIdog
- Grok 4.5 launched: what xAI's benchmarks show vs Opus 4.8 - roo.beehiiv
- SpaceX to acquire Cursor for $60B - TechCrunch
Last updated
✓ Last verified July 9, 2026
