Grok 4.5: Coding Benchmarks, 80 TPS, and Token Efficiency

SpaceXAI launched Grok 4.5 on July 16, 2026 for coding, agentic work, and knowledge tasks. The official announcement reports 83.3% on Terminal-Bench 2.1, 64.7% on SWE-bench Pro, and a 29.0% pass-at-1 resolution rate on SWE Marathon. The model is available through the SpaceXAI API, Grok Build, and Cursor.

The release also makes an unusually concrete efficiency claim: 80 generated tokens per second and an average of 15,954 output tokens per SWE-bench Pro task, versus 67,020 for Claude Opus 4.8 at its cited maximum setting. That is 4.2 times fewer output tokens in the vendor's run. Capability and efficiency should still be reproduced together because a shorter trace that fails more often is not cheaper work.

What launched on July 16

Grok 4.5 is a hosted frontier model; the announcement does not offer downloadable weights. SpaceXAI says it was trained alongside Cursor and on data spanning software, science, engineering, and mathematics. Reinforcement-learning rollouts covered hundreds of thousands of multi-step tasks and ran asynchronously for many hours across tens of thousands of NVIDIA GB300 GPUs.

The product positioning combines engineering agents with office work. Grok 4.5 is the default model in Grok Build and is shown producing spreadsheets, presentations, documents, and complete web applications. Those demonstrations help define intended workloads, but the quantitative evidence is strongest for code and terminal agents.

The official model documentation identifies grok-4.5 as the API model and recommends a conversation-level prompt cache key. Teams should pin the exact model identity and interface because a product default can change without preserving evaluation comparability.

First-party benchmark table

SpaceXAI publishes these values and names each evaluation harness:

Competitor rows in the launch charts come from developers' system cards or benchmark leaderboards. That is useful context but not a controlled universal ranking. Grok's scores alone are the most portable facts; cross-model comparisons must retain effort, harness, tool, context, and date.

DeepSWE results depend on the harness

DeepSWE 1.0 gives Grok 4.5 a 62.0% score. SpaceXAI says the evaluation was created by Datacurve, used each provider's harness, and was run by Artificial Analysis. DeepSWE 1.1 instead uses mini-swe-agent and gives Grok 53.0%. The drop does not imply regression: these are different benchmark versions and agent arrangements.

The release chart places Fable and GPT-5.5 ahead on both DeepSWE versions, while Grok remains competitive. That makes the vendor's claim more credible than a cherry-picked “best at everything” narrative. It also shows why model and scaffold should be selected as a system.

For internal evaluation, keep one neutral harness beside each vendor-native harness. The neutral run tests portability; the native run tests attainable product value. Our software-engineering agent guide explains how to record patch validity, test coverage, regressions, and reviewer effort rather than only resolution rate.

SWE Marathon and long-horizon work

SWE Marathon's 29.0% pass-at-1 result is the one chart where Grok 4.5 leads the cited set: Opus 4.8 is shown at 26.0%, Fable at 24.0%, and Opus 4.7 at 16.0%. Long-running tasks expose planning drift, context pressure, dependency mistakes, and unproductive retries that smaller issue benchmarks can hide.

Pass-at-1 is operationally relevant because it does not grant repeated sampling to find one lucky solution. But the release page does not make every cost and timeout directly comparable across providers. A local reproduction should set a wall-clock limit, tool-call ceiling, compute budget, and identical repository snapshot.

Record partial progress as well as final pass. An agent that finds the right subsystem and writes useful tests before timing out may be valuable in an assisted workflow, while an opaque patch that happens to pass can still create maintenance risk.

Terminal and repository benchmarks

On Terminal-Bench 2.1, Grok 4.5 scores 83.3%, essentially level with the cited GPT-5.5 value of 83.4 and Fable's 84.3. Terminal tasks reward environment inspection, package and process control, and verification. The public Terminal-Bench project is the right place to inspect task and harness evolution rather than treating a versionless score as permanent.

SWE-bench Pro is 64.7% for Grok. SpaceXAI's chart puts Fable at 80.4, Opus 4.8 at 69.2, Opus 4.7 at 64.3, GLM 5.2 at 62.1, and GPT-5.5 at 58.6 under their cited configurations. The SWE-bench project documents the broader evaluation family, but Pro runs may use commercial subsets and distinct scaffolds.

Neither number replaces a private repository trial. Include monorepos, generated code, migration scripts, weak tests, secret boundaries, and failing dependencies. Require the model to explain evidence and identify uncertainty before merging.

Speed and token efficiency

SpaceXAI reports 80 TPS and 15,954 average output tokens per SWE-bench Pro task. The cited Opus 4.8 maximum configuration used 67,020, making Grok's trace about 4.2 times smaller. The page also describes roughly twice the token efficiency of comparable leading models more generally.

Output tokens are only part of agent cost. Input context, cached reads, tool execution, container time, retries, and human correction can exceed generation charges. A compact trace can also hide additional server-side reasoning depending on billing semantics. Compare accepted tasks, not raw calls.

At launch pricing, one million input tokens costs $2 and one million output tokens $6. Build a distribution of cost per accepted task and include failures. The inference latency guide provides a template for separating time-to-first-token, generation speed, tool waiting, and end-to-end completion.

Training claims and evidence limits

The announcement says training used tens of thousands of GB300 GPUs, aggressive data deduplication and quality scoring, and asynchronous RL across hundreds of thousands of tasks. These statements explain the intended sources of engineering performance but do not provide a reproducible training recipe or weights.

The model is therefore auditable through API behavior and published evaluation, not by inspecting parameters. Enterprises should archive prompts, outputs, model names, dates, prices, safety settings, and evaluation harness commits. Repeat a canary set after any observed behavior change.

The absence of an attached Grok 4.5 system card in the launch material limits safety comparison. Treat capability scores as one side of acceptance; separately test prompt injection, secret handling, destructive tool use, deception, and refusal calibration under the exact agent permissions.

Production acceptance plan

Start with read-only repository analysis and bounded terminal containers. Move to patch generation only after the model consistently scopes changes, runs relevant tests, and reports failures honestly. Sending, merging, deployment, credential changes, and deletion should remain separately authorized actions.

Evaluate three tiers: short fixes, multi-file feature work, and day-scale maintenance. For each, measure pass rate, regression rate, elapsed time, total token and tool cost, reviewer minutes, unsafe attempts, and recovery from interrupted sessions. Keep a smaller or cheaper model as a baseline.

Prompt caching is useful for long conversations, but cache identity must not cross users, repositories, or permission contexts. Redact secrets before sending context and expire project caches when access changes.

Verdict

Grok 4.5 is a significant coding-agent release because it combines strong terminal and repository results with explicit speed, price, and output-token figures. The most notable balance is 83.3% on Terminal-Bench 2.1 and 64.7% on SWE-bench Pro at a claimed 80 TPS, not a claim of winning every benchmark.

The right next step is a system-level evaluation against the organization's repositories and tools. Preserve the published harness distinctions, price all failed attempts, and keep material actions behind approval. If the 15,954-token average survives that controlled test without lowering resolution or safety, Grok 4.5 could be unusually efficient for long coding workflows.

Source notes — reviewed 2026

• SpaceXAI's Grok 4.5 announcement provides release date, benchmark values, speed, token use, training claims, and pricing.
• The official Grok 4.5 developer documentation documents the API model and integration guidance.
• The SWE-bench project documents the repository-resolution benchmark family.
• The Terminal-Bench project documents terminal-agent evaluation and evolving benchmark versions.