What Is Grok 4.5? Capabilities, Benchmarks, and Adoption Decisions

As of August 9, 2026, Grok 4.5 is an officially released SpaceXAI model, not an unverified preview name. The launch announcement positions it for coding, agentic tasks, and knowledge work. The developer guide confirms the API model ID grok-4.5 and support through the Responses API and Chat Completions.

The useful engineering question is not whether it is “the strongest.” It is which capabilities the public evidence supports, which figures remain vendor evaluations, and whether quality, latency, cost, and risk improve after the model is placed inside your own coding-agent harness.

Verified Product Identity and Availability

SpaceXAI’s public documentation provides the following operational facts:

Dimension	Official information as of 2026-08-09	Adoption implication
API	Model ID grok-4.5; Responses API and Chat Completions	Existing OpenAI-compatible clients can support a controlled trial
Input and output	Text and image input, text output; 500K context window	Large repositories still need context selection; capacity is not retrieval quality
Reasoning and tools	Low, medium, or high reasoning; function calling, web search, X search, and code execution	Tool permissions and sandboxing directly affect success and risk
Knowledge and recency	Knowledge cutoff February 1, 2026; current information requires search tools	Model memory is not a source for current dependencies or security notices
Pricing	US$2 per million input tokens, US$0.30 cached input, and US$6 output	Measure retries, tool results, and compaction across the full task
Product surfaces	xAI API, Grok Build, Cursor on all plans, and Office add-ins	Availability does not imply identical permissions, data policy, or latency across surfaces

The announcement says Grok 4.5 was “trained alongside Cursor,” but it does not explain data exchange, training responsibility, or the scope of joint development. Describing this as “co-trained by SpaceXAI and Cursor” would go beyond the public evidence.

Public Training Disclosures Are Not a Full Architecture

xAI says Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs on material spanning coding, science, engineering, and math, with deduplication, quality scoring, and domain-focused selection. Its reinforcement-learning stage covered hundreds of thousands of multi-step software-engineering and technical tasks with automated and model-based grading. An asynchronous training stack allowed long agent rollouts to proceed while learning continued elsewhere.

The conservative architecture reading is that Grok 4.5’s engineering performance comes from a combination of model training, inference budget, tool interfaces, and execution harness. The cited public pages do not support claims about parameter count, network topology, the full training-data mixture, or detailed safety training because they do not disclose those properties.

For long tasks, the developer guide recommends prompt_cache_key to route a conversation for reliable cache hits and context compaction for long agent loops. This reinforces the lesson from Bloss0m’s long-running agent harness guide: a larger context window does not replace explicit state handoff, tests, and recovery points.

What the Benchmarks Do—and Do Not—Show

xAI’s launch post reports five software-engineering evaluations:

Evaluation	Reported Grok 4.5 result	How to read it
DeepSWE 1.0	62.0%	Models used their providers’ harnesses, so the execution environment was not uniform
DeepSWE 1.1	53%	Datacurve ran it with the mini-swe-agent harness
SWE Marathon	29.0% pass@1	It signals longer-task ability but is still not your codebase
Terminal Bench 2.1	83.3%	Controlled terminal success does not establish production operational safety
SWE Bench Pro	64.7% resolve rate	Useful as a relative signal, not a result that generalizes to every language and repository

The same announcement claims service speed of roughly 80 tokens per second. It also reports an average of 15,954 output tokens on SWE Bench Pro, about 4.2 times fewer than Opus 4.8 (max) at 67,020. That is a testable cost hypothesis, but it remains xAI’s comparison. Reasoning settings, harness behavior, retries, caching, and tool output can all change end-to-end cost.

The chart also does not show Grok 4.5 leading every category. In xAI’s own figures, other models score higher on DeepSWE 1.0, DeepSWE 1.1, Terminal Bench 2.1, and SWE Bench Pro. “Dominates every benchmark” would therefore misstate the source.

Four Engineering Decisions Before Adoption

1. Build an acceptance set from your workload

Cover bug fixes, cross-file changes, test creation, dependency migrations, and failure recovery. In addition to pass rate, record unnecessary diffs, test validity, human correction time, and rollbacks. Bloss0m’s review of agentic-coding training and evaluation boundaries offers a useful starting point for replayable comparisons.

2. Measure model cost separately from agent cost

Track input, cached input, output, search and code-tool usage, wall-clock latency, and retries. A 500K context window is not a reason to fill every request. Prompt caching only helps when prefixes remain stable enough to reuse.

3. Define tool permissions before optimizing success rate

When a coding model can call a shell, search, or external systems, default to least privilege, isolate secrets, constrain network and filesystem access, and require human approval for high-impact commands. The AI Agent architecture, evaluation, and enterprise guide covers the wider control plane.

4. Choose a version-change policy

grok-4.5-latest suits exploratory environments that accept ongoing changes. For reproducible CI or regulated workflows, confirm dated-version availability, retirement policy, and fallback behavior with the provider. A stable model name does not guarantee immutable backend behavior.

Practical Assessment

Grok 4.5 has a formal API, explicit pricing, long context, and several engineering benchmarks, which is enough to place it on a candidate list. The public evidence is not enough to call it an independently proven “strongest coding model.” The responsible next step is a small bake-off with a fixed harness: identical issues, permissions, tests, and cost accounting, evaluated on whether it actually reduces completion time and human correction.

Primary Sources

• SpaceXAI: Introducing Grok 4.5
• SpaceXAI developer guide: Grok 4.5
• SpaceXAI model detail: grok-4.5
• Official SpaceXAI Grok Build repository