Grok 4.5: 500K Context, 64.7% SWE-bench Pro (2026)
Grok is xAI's July 2026 coding model, generating output at about 80 tokens per second, with pricing well below Claude's flagship for high-volume agentic work.
This model suits teams running high-volume agentic coding where speed and per-task cost outweigh chasing the very top accuracy score. It resolves the average coding task in about 15,954 output tokens, far fewer than pricier rival models use for similar work.
Grok 4.5, launched by xAI on July 8, 2026, is a coding and agentic-workflow model built for high-volume tool-calling and CI automation. It scored 83.3% on Terminal-Bench 2.1, trading top-end reasoning accuracy for markedly faster, cheaper output than rival flagship models.
Provider: xAI · Family: Grok 4
Context window: 500,000 tokens
Input modalities: text, image, tool-calls · Output: text, tool-calls
About Grok 4.5
Grok 4.5 is xAI's coding and agentic-workflow model, trained jointly with Cursor and released to private beta customers on July 8, 2026 before opening to the public a day later through Grok Build, Cursor, and the xAI API console. It is the first Grok release xAI built specifically for software engineering and long-horizon agent loops rather than general chat. xAI trained it on tens of thousands of Nvidia GB300 GPUs, with a reinforcement learning stage that covered hundreds of thousands of largely software-engineering tasks, on top of a foundation architecture referred to in leaked training traces as V9 (parameter count unconfirmed). At launch xAI published agentic and coding evaluations rather than the classic academic benchmark suite. Grok 4.5 scored 64.7% on SWE-bench Pro, ahead of GPT-5.5's 58.6% on the same measure but behind Claude Opus 4.8 (69.2%) and Claude Fable 5 (80.4%). It posted an 83.3% resolution rate on Terminal-Bench 2.1 and 29.0% on SWE Marathon, a long-horizon agentic-coding test. xAI did not disclose GPQA Diamond, AIME 2025, MMLU-Pro, or ARC-AGI 2 scores at launch, so how it stacks up on pure academic reasoning is unverified. The model ships a 500,000-token context window, the smallest of the three current frontier models, though still enough for most repository-scale coding tasks. Reasoning effort is configurable across low, medium, and high, with high as the default, and xAI added context compaction to keep long agent sessions inside the window without losing task state. Grok 4.5 accepts text and image input and returns text only, so it has no native audio or video modality. It supports function and tool calling, structured JSON output, web search, X search, code execution, and document search across collections, covering most of what an agentic coding pipeline needs without a separate orchestration layer. On speed, xAI's own reporting and independent testers put Grok 4.5 at roughly 80 tokens per second, faster than Claude Opus 4.8 in head-to-head measurements, and the model resolves coding-agent tasks using an average of 15,954 output tokens versus Opus 4.8's 67,020, a gap that compounds with its lower per-token price. Pricing itself sits well under every current frontier rival, which is the core of xAI's pitch for teams running constant, high-volume agent workloads (exact rates are in the pricing section below). The model is available through the xAI API console, Grok Build (xAI's separately released coding-agent CLI, licensed Apache 2.0, though that release covers only the agent runtime and not the Grok 4.5 weights), Cursor, Microsoft Office add-ins, and model gateways including OpenRouter, Vercel AI Gateway, Cloudflare Workers AI, Snowflake Cortex, and Databricks Mosaic AI. Grok 4.5 itself remains closed-weight and API-only; nothing about the underlying model has been open-sourced. Knowledge cutoff is February 1, 2026, and xAI describes the training mix as spanning code, science, engineering, and math, filtered by deduplication and quality scoring rather than raw volume. Unlike xAI's Grok 4.1 (November 17, 2025) and Grok 4.20 (April 7, 2026) releases, both of which shipped with a public system card at data.x.ai, Grok 4.5 launched without one, leaving its documented safety evaluation thinner than its two immediate predecessors. The combination of low cost and high token efficiency makes Grok 4.5 the strongest fit for teams running constant, high-volume agentic coding or CI automation where cost per resolved task outweighs squeezing out the last few points of accuracy. Teams that need the single highest raw accuracy on hard SWE-bench Pro tasks, the largest available context window, or a published system card before deployment are better served by Claude Opus 4.8, Claude Fable 5, or GPT-5.5 instead.
Pricing
Standard tier: $2.00 per 1M input tokens, $6.00 per 1M output, and $0.50 per 1M cached input, all for prompts up to 200,000 tokens. Cross that threshold and rates roughly double: $4.00 input, $1.00 cached, $12.00 output per 1M, still far below Claude Opus 4.8's $5/$25 per-million rate.
Key Features
- Large Context Window: Sized for repository-scale codebases and long agent sessions, with compaction to preserve task state as sessions near the limit.
- Prompt Caching: Cached input tokens cost a fraction of standard input pricing, cutting costs for repeat-context agent loops.
- Configurable Reasoning Effort: Low, medium, or high reasoning modes let callers trade cost and latency against accuracy per request.
- Native Tool and Code Execution: Built-in function calling, structured JSON output, web and X search, and code execution cover most agentic-coding needs without extra orchestration.
- High Token Efficiency: Resolves coding-agent tasks using far fewer output tokens than pricier rival models, per independent testing.
Pros
- Cheapest frontier-class pricing among current flagship models, undercutting every major rival on both input and output tokens.
- Fastest output speed among current frontier models in head-to-head measurements.
- Outperforms OpenAI's competing flagship on the SWE-bench Pro coding benchmark, the one head-to-head score xAI published at launch.
Cons
- GPQA Diamond, AIME 2025, MMLU-Pro, and ARC-AGI 2 scores were absent from the launch materials, leaving pure-reasoning performance versus rivals unverified.
- Smallest context window of the three current frontier models.
- Launched without a public system card, unlike xAI's own two most recent prior releases.
Benchmarks
- swe marathon: 29
- swe bench pro: 64.7
- terminal bench 2 1: 83.3
- artificial analysis speed tokens per sec: 80
Frequently Asked Questions
How much does Grok 4.5 cost per 1M tokens?
Prompts up to 200,000 tokens run at $2.00 per 1M input, $0.50 per 1M cached input, and $6.00 per 1M output. Longer prompts shift to a pricier tier: $4.00 input, $1.00 cached, and $12.00 output per 1M tokens.
How does Grok 4.5 compare on benchmarks vs GPT-5.5?
It scores 64.7% on SWE-bench Pro versus 58.6% for GPT-5.5, and generates output faster while using fewer tokens per resolved task. Neither model published GPQA, AIME, or MMLU-Pro scores at launch, so their pure-reasoning gap is unconfirmed.
Is Grok 4.5 open source or proprietary?
It is proprietary and API-only; no model weights have been released. xAI separately open-sourced Grok Build, its Apache 2.0 coding-agent CLI, but that release does not include the model itself.
Does Grok 4.5 train on user data?
xAI has not published a system card or data-retention policy specific to this model as of its July 2026 launch, unlike its two immediate predecessors, which both shipped with public system cards. Enterprise buyers should request retention terms directly from xAI before deployment.
Who is Grok 4.5 best for and who should avoid it?
It fits teams running high-volume agentic coding or CI automation where per-task cost and speed matter most, given its fast output and low token usage versus pricier rivals. Teams needing the top SWE-bench Pro accuracy, the largest context window, or a published system card should pick a rival flagship instead.