Grok 4.5 Review: Coding, Pricing, Benchmarks, and Limits

This Grok 4.5 review examines SpaceXAI's coding-focused model without treating launch claims or leaderboard positions as a universal verdict. Released publicly in July 2026, Grok 4.5 combines a 500,000-token context window, configurable reasoning, image input, function calling, structured output, and distribution through the API, Grok Build, Cursor, and several model gateways.

The headline is price-to-performance. API rates are $2 per million input tokens and $6 per million output tokens, while independent evaluation places the model near the intelligence frontier and shows particularly strong agentic coding results.

Grok 4.5 Review: Quick Verdict

Grok 4.5 is a credible candidate for coding agents, technical research, and tool-heavy workflows where total task cost matters. It offers near-frontier benchmark performance at lower list prices than several flagship alternatives, and SpaceXAI says it was trained alongside Cursor using supplemental anonymized workflow data.

Its strengths are not uniform. Independent Artificial Analysis results support strong overall intelligence and coding-agent performance, but the model does not lead every reasoning or software benchmark. SpaceXAI's own tables also show it behind competing models on some tests.

The practical verdict is to test Grok 4.5 as a cost-efficient agent model, not to migrate on price alone. Measure accepted-task rate, latency, output tokens, tool calls, retries, and reviewer time. Keep human approval around production changes and other consequential actions.

What Is Grok 4.5?

Grok 4.5 is SpaceXAI's proprietary model for coding, engineering, agentic tasks, and knowledge work. Its API identifier is grok-4.5, with grok-4.5-latest available as an alias. The model accepts text and images and returns text.

The current developer documentation lists a 500,000-token context window, configurable reasoning, function calling, and structured output. SpaceXAI's model directory gives it a February 1, 2026 knowledge cutoff, while the model card describes a January 2026 pretraining cutoff. Either way, current facts require web search, connected sources, or another retrieval layer.

Access extends beyond a raw API. Grok 4.5 is the default model in Grok Build, SpaceXAI's terminal coding agent, and is available in Cursor on all plan tiers. The model card also lists Office add-ins and several model gateways. Verify product and regional availability before planning a rollout.

Grok 4.5 Coding and Agent Performance

What SpaceXAI Reports

SpaceXAI emphasizes repository work, terminal use, long-horizon engineering, and deliverable creation. Its launch material reports 62% on DeepSWE 1.0, 53% on DeepSWE 1.1, 83.3% on Terminal-Bench 2.1, and 64.7% on SWE-Bench Pro. The model also scored 29% on SWE-Marathon in the vendor's published comparison.

Grok 4.5 trails Fable 5 and GPT-5.5 on both DeepSWE versions, nearly ties GPT-5.5 on Terminal-Bench 2.1, and leads the listed models on SWE-Marathon. Harnesses, reasoning settings, token budgets, graders, and task sets can change the order.

SpaceXAI also claims high token efficiency. On its SWE-Bench Pro comparison, Grok 4.5 averaged 15,954 output tokens per task versus 67,020 for Opus 4.8 at maximum effort. That is a vendor-selected comparison, but it highlights the right production metric: cost per accepted result can matter more than cost per token.

What Independent Testing Adds

Artificial Analysis scored Grok 4.5 at 54 on its Intelligence Index. Its July release analysis placed the model near the leading proprietary systems and reported a score of 76 on the Coding Agent Index when run in the Grok Build harness, roughly level with GPT-5.5 in Codex in that evaluation and below Fable 5 in Claude Code.

The durable signal is that Grok 4.5 combines strong agentic performance with relatively concise output and lower API rates than many models in its capability class.

Do not interpret that as proof that Grok 4.5 will write the most maintainable patch, follow your architecture, or catch the right regression. Agent results depend on the model plus the harness, repository instructions, tools, environment, and review loop. SD's Claude Code vs Codex comparison explains why the surrounding workflow can change the result as much as the base model.

Grok 4.5 Pricing and Context Limits

The standard SpaceXAI API price is $2 per million input tokens and $6 per million output tokens. SuperGrok, the consumer subscription that includes Grok 4.5 and higher limits, is listed at $30 per month. Consumer subscriptions and API billing are separate products.

The 500,000-token context window can hold large repositories, document sets, or extended agent traces. Requests that exceed 200,000 tokens use different higher-context pricing, according to the dedicated model page. SpaceXAI does not show the higher rate in the rendered summary, so teams should verify the console price before sending very large prompts.

Large context is capacity, not proof of reliable recall. Retrieval, context compaction, and evidence requirements still matter, and targeted retrieval can be cheaper than repeatedly sending an entire repository.

For comparison, GPT-5.6 Sol currently costs $5 per million input tokens and $30 per million output tokens, while Claude Sonnet 5 has introductory pricing of $2 and $10 through August 31, 2026. Grok 4.5 is cheaper than both on output list price, but list price excludes differences in tokenization, reasoning effort, retries, tool fees, and reviewer cleanup. SD's GPT-5.6 review covers the OpenAI family in more detail.

Grok 4.5 Features That Matter

Configurable reasoning lets developers trade response time and token use for deeper work. Start with the default, then compare lower and higher settings on representative tasks rather than assuming maximum effort is always economical.

Function calling connects the model to application-defined tools, while structured output helps enforce a response schema. SpaceXAI also offers server-side web search, X search, code execution, and collections search. Tool charges can sit on top of token charges, and tools expand the impact of a wrong decision.

The February knowledge cutoff is especially relevant for a model marketed around engineering. Package versions, security advisories, product documentation, and cloud behavior change quickly. Enable current retrieval and require citations when freshness affects the answer.

Grok 4.5 vs GPT-5.6 and Claude

Choose Grok 4.5 when strong coding-agent performance and low output cost are the priorities. Its 500,000-token window is substantial, and distribution through Grok Build and Cursor makes it easy to test in real repositories.

Choose GPT-5.6 when your workflow benefits from the OpenAI Responses API, Programmatic Tool Calling, multi-agent beta, or the option to route across Sol, Terra, and Luna. GPT-5.6 Sol has a larger 1.05-million-token window but a much higher flagship price.

Choose Claude Sonnet 5 when Claude Code, Anthropic's tool ecosystem, or its one-million-token context better fits the workflow. Sonnet 5's introductory input price matches Grok 4.5, but its output rate is higher. That price comparison can change after the promotional period.

There is no clean winner across coding, office work, research, and customer-facing agents. Test the same straightforward, ambiguous, long-context, and tool-error cases inside the harness you will use.

Safety, Reliability, and Governance

SpaceXAI's 29-page model card is more useful than a general safety promise. It documents coding, engineering, knowledge-work, cyber, biology, jailbreak, refusal, child-safety, mental-health, and behavioral evaluations. It also states that Grok 4.5 is not intended for autonomous high-stakes decisions in medicine, law, finance, or safety-critical systems without human oversight and domain-expert validation.

Most results in a vendor model card are produced, selected, or compiled by the vendor. They are valuable evidence, not a warranty. Applications still need source checks, least-privilege credentials, isolated execution, logs, spending limits, and approval gates.

For coding agents, require tests and human review before merging. A plausible patch can still introduce security, maintenance, or licensing problems. SD's AI agent security guide provides a broader control framework.

Who Should Use Grok 4.5?

Grok 4.5 is worth testing for teams building coding agents, technical research tools, engineering assistants, or structured workflows with frequent tool calls. It may also suit developers already using Cursor or teams that want a lower-cost alternative to flagship model APIs.

It is less compelling when a smaller model already clears the quality bar, when a workflow needs more than 500,000 tokens, or when organizational policy does not support the vendor, region, or data path. Teams without an evaluation set should create one before changing production traffic.

A safe rollout moves from read-only analysis to draft changes, then sandboxed execution, and finally narrowly scoped actions. At each stage, compare task success and failure cost—not just benchmark rank.

Conclusion

This Grok 4.5 review finds a serious coding and agent model with an attractive price, a large context window, broad distribution, and credible independent benchmark support. Its strongest case is not that it wins every leaderboard. It is that it can deliver near-frontier performance with concise output and relatively low API rates.

The caveats are equally practical. Vendor benchmarks vary by setup, independent ranks move, long context has pricing and reliability costs, and tool access increases risk. Evaluate Grok 4.5 on real tasks, measure cost per accepted result, and keep consequential actions behind human review. That evidence—not launch-day positioning—should decide whether it earns production traffic.