Grok 4.5 Review: 500K Context, Speed, Coding & Agents
After Grok 4.5 launched, I read xAI's announcement and developer documentation, then compared them with the first discussions on X and Reddit.
The short version is simple: Grok 4.5 is not a minor chat-model refresh. It is xAI's flagship model for coding, agentic tasks, and knowledge work.
The questions developers actually care about are more practical:
- Can Grok 4.5 compete with Claude, GPT, and GLM on real repositories?
- Do the claimed 80 TPS and lower token usage matter in long agent runs?
- Is the 500K context window useful for large codebases?
- Is the official price of $2 per million input tokens and $6 per million output tokens genuinely competitive?
- Why do some users describe a major upgrade while others complain about access and usage limits?
This review separates verified product facts from early community impressions and explains where Grok 4.5 may fit in a production model stack.
Quick verdict
- Grok 4.5 launched on July 8, 2026 under the model ID
grok-4.5. - xAI positions it for coding, agentic tasks, and knowledge work.
- It supports a 500K-token context window and has a February 1, 2026 knowledge cutoff.
- xAI reports serving speed of approximately 80 tokens per second.
- Official API pricing is $2 per million input tokens and $6 per million output tokens.
- Reasoning effort is configurable as
low,medium, orhigh, withhighas the default. - Its strongest use cases appear to be repository-level engineering, terminal tasks, long agent workflows, office files, and constrained knowledge work.
My conclusion: Grok 4.5 belongs in the evaluation pool for any team building coding agents or multi-model infrastructure. It is too early, however, to call it an uncontested winner based only on vendor benchmarks.
Grok 4.5 specifications
| Item | Grok 4.5 |
|---|---|
| Model ID | grok-4.5 |
| Architecture | Mixture of Experts (MoE) |
| Main focus | Coding, agents, knowledge work |
| Context window | 500K tokens |
| Serving speed | Approximately 80 TPS, according to xAI |
| Reasoning effort | low / medium / high |
| Official input price | $2 / 1M tokens |
| Official output price | $6 / 1M tokens |
| Knowledge cutoff | February 1, 2026 |
| Real-time information | Requires Web Search or X Search tools |
The product strategy becomes clearer when these specifications are considered together. A large context window holds the repository and supporting documents, high throughput reduces waiting, and token efficiency lowers the cost of multi-step execution.
That combination is designed for agent workflows, not merely for winning short-form chat comparisons.
Coding performance: strong numbers, with an important caveat
xAI published results across DeepSWE, SWE Marathon, Terminal Bench 2.1, and SWE Bench Pro. Representative scores include:
- 29.0% on SWE Marathon
- 83.3% on Terminal Bench 2.1
- 64.7% on SWE Bench Pro
- 62.0% on DeepSWE 1.0
These benchmarks align with the model's intended workload: understanding real repositories, operating a terminal, editing multiple files, running tests, and maintaining goals across long execution chains.
There is also a caveat developers should not ignore. Cursor disclosed that an earlier snapshot of its codebase was accidentally present in Grok 4.5 training data, giving the model an advantage on CursorBench. The exact impact is unclear, and the disclosure does not invalidate every other benchmark, but it makes one principle especially important:
Evaluate Grok 4.5 on your own repositories, tests, and cost constraints instead of treating a leaderboard as a deployment decision.
Why 80 TPS may matter more than another benchmark win
For a short answer, 80 TPS simply feels responsive. For an agent, the speed advantage compounds across dozens of steps: reading files, calling tools, editing code, running tests, inspecting failures, and trying again.
A modest latency improvement in each step can reduce the total task time by several minutes.
xAI also reports that Grok 4.5 used an average of 15,954 output tokens per SWE Bench Pro task in its comparison, versus 67,020 for Opus 4.8 max, or roughly 4.2 times fewer output tokens.
That is a vendor-provided comparison and should not be generalized to every workload. Still, it points to the right cost metric. Teams should compare:
- total tokens per completed task
- wall-clock completion time
- number of retries
- test pass rate
- human review and repair time
The cheapest token is not always the cheapest completed task.
What can you do with a 500K context window?
Five hundred thousand tokens is sufficient for many substantial engineering and knowledge workflows, including:
- cross-module analysis in medium and large repositories
- combined review of pull requests, issues, logs, and technical documentation
- contracts, research libraries, and enterprise knowledge bases
- long-running agents that preserve execution history
- multi-file Word, Excel, and PowerPoint workflows
A large window does not guarantee perfect recall at every position, and it does not justify dumping an entire repository into every request. A better workflow is to build a dependency map, retrieve relevant files, preserve concise decision summaries, and checkpoint long tasks between phases.
What Reddit users are saying
Early Reddit reactions are mixed, which makes them more useful than a launch announcement.
Positive feedback clusters around three areas.
First, users report better behavior under complex instructions. One discussion praised Grok 4.5 for handling ambiguous premises and demanding writing prompts without collapsing into an obvious, superficial answer.
Second, developers frequently mention speed. In communities comparing Grok with Sonnet, Opus, GPT, and GLM, throughput and potential cost per completed task are among the most attractive features.
Third, several users describe a clear gap between the API or Grok Build experience and the less transparent consumer app. This matters because a vague model selector makes it difficult to know whether a comparison is actually testing Grok 4.5.
The negative feedback is equally consistent.
Users have complained that launch access was confusing across the web app, mobile app, Heavy, Expert, Grok Build, and API. Asking the model which version it is running is not reliable because it may infer an answer from web content.
Usage limits are another source of frustration. One user reported that a Grok Build task stopped around 105K of a possible 500K context tokens after only a short period. Others objected to weekly limits because they make long projects difficult to predict.
The practical takeaway is that model capability and product usability are separate layers. A strong model can still provide a poor experience if limits, routing, and recovery behavior are unclear.
What the X discussion adds
Discussion on X has focused on the Cursor collaboration, one-prompt application demos, model competition, and xAI's rapid release cadence.
Those topics are useful for discovery, but production buyers should track less glamorous metrics:
- fixed model aliases and version stability
- peak-hour latency and error rates
- token use on long tasks
- tool-call success rates
- quality differences across reasoning levels
These measurements are more valuable than viral demos when the model is part of a customer-facing service.
Is Grok 4.5 API pricing competitive?
At $2 per million input tokens and $6 per million output tokens, the list price is competitive for a frontier coding model. The final cost, however, has at least four layers: repeated long-context input, reasoning and code output, success rate, and platform overhead.
If your team tests Grok alongside GPT, Claude, GLM, DeepSeek, and other models, a unified model API platform such as llm-agent can simplify the comparison. A common API key, consistent request patterns, usage tracking, and model switching reduce the operational work required to evaluate another provider.
See model pricing and availability, buy API tokens, and the integration tutorials. Current model availability, regional support, and platform pricing should always be verified on the live pricing page.
Who should test Grok 4.5?
Grok 4.5 is particularly relevant for:
- engineering teams using agents on medium or large repositories
- independent developers building complete prototypes and applications
- knowledge workers producing complex spreadsheets, documents, and presentations
- model gateways reducing dependency on a single provider
- businesses comparing the task-level cost of Claude, GPT, Grok, and Chinese coding models
It is less suitable for an immediate, untested migration where output consistency is critical, for users who cannot verify the actual model behind a subscription interface, or for projects that require immediate EU availability and specific compliance guarantees.
Final assessment
The most interesting aspect of Grok 4.5 is not a single benchmark score. It is the combination of flagship-level engineering capability, fast-model throughput, and a focus on reducing the number of tokens required to finish a task.
Official data and early community reports both suggest a meaningful improvement for complex instructions, terminal work, long execution chains, and knowledge workflows.
There are still reasons for caution: confusing launch access, unpredictable subscription limits, a disclosed training-data caveat around CursorBench, and limited public evidence from long-running production projects.
The honest verdict is therefore straightforward: if you build coding or agent products, test Grok 4.5 now on real work. If you are considering a full migration based only on launch benchmarks, wait for your own evaluation results.
FAQ
When was Grok 4.5 released?
Grok 4.5 was officially released on July 8, 2026 and made available through the xAI API, Grok Build, and Cursor-related access paths.
What is the Grok 4.5 context window?
xAI's model documentation lists a 500K-token context window.
How much does the Grok 4.5 API cost?
The official standard price is $2 per million input tokens and $6 per million output tokens. Third-party and unified API platforms may use different pricing.
Is Grok 4.5 good for coding?
Coding and agentic tasks are its primary positioning. Its benchmarks and early user reports suggest it is most useful for repository work, terminal operations, and multi-step development rather than isolated code snippets.
Can Grok 4.5 search X and the live web?
Yes, when Web Search or X Search tools are enabled. The model's internal knowledge cutoff is February 1, 2026.
Is Grok 4.5 better than Claude, GPT, or GLM?
No model wins every workload. Grok 4.5 is attractive for speed, coding, agent execution, and list pricing, while competitors may lead on particular repositories, design tasks, ecosystems, or long-run reliability. Compare task success, total tokens, time, and retries on the same private benchmark.