Grok 4.5 Is Live on the API: Token Efficiency Math That Changes Your Coding Agent Routing Policy

xAI launched Grok 4.5 today — its smartest model yet, tuned specifically for coding, agentic workflows, and knowledge work. The model is available immediately on the xAI API at $2 per million input tokens and $6 per million output tokens, with a $0.50/M cached input tier. It serves at 80 tokens per second and carries a 500K-token context window.

Those four numbers together — price, output speed, context depth, and token efficiency — make Grok 4.5 a routing decision, not just a model upgrade.

What happened

xAI released Grok 4.5 on July 17, 2026. The model was trained across tens of thousands of NVIDIA GB300 GPUs, with a focus on high-signal data curation and reinforcement learning over multi-step software engineering tasks. Unusually, Grok 4.5 was developed in close collaboration with Cursor; Cursor users on all plans can access it immediately. It also becomes the default model in Grok Build, xAI's coding agent.

Benchmark highlights from xAI's published evaluations:

• SWE Marathon (pass@1): Grok 4.5 at 29.0%, ahead of Opus 4.8 (26.0%) and Claude Fable (24.0%)
• SWE Bench Pro: Grok 4.5 64.7%, trailing Opus 4.8 (69.2%) but ahead of GPT 5.5 (58.6%)
• Terminal Bench 2.1: Grok 4.5 at 83.3%, within 1% of GPT 5.5 and Claude Fable

The headline operator finding: Grok 4.5 resolves SWE Bench Pro tasks with 15,954 average output tokens versus Opus 4.8's 67,020 — roughly 4.2× fewer tokens per task at comparable resolution rates.

Why it matters for AI engineering teams

Output token count is the dominant cost driver in agentic coding loops. A model that resolves issues in fewer tokens — without sacrificing resolution rate — compresses cost-per-task linearly. When you plug the numbers in:

• Opus 4.8 (max): ~67,000 output tokens × $15/1M = $1.00 per average SWE task
• Grok 4.5: ~16,000 output tokens × $6/1M = $0.096 per average SWE task

That's roughly a 10× cost difference per resolved coding task, not accounting for input tokens. Even if Grok 4.5 falls short on your specific task class, the efficiency gap is large enough that it belongs in your routing evaluation.

At 80 TPS, Grok 4.5 also avoids the latency penalty that large reasoning models typically impose. It occupies the "large and fast" slot that's rarely available — models at this performance tier usually require explicit tradeoffs between speed and quality.

The router/operator angle

For teams running an AI gateway or model router, Grok 4.5 presents three immediate routing considerations:

1. Tier placement. Grok 4.5 sits between a standard completion model and a premium reasoning model on the cost curve, but closer to a fast model on latency. Teams with tiered routing policies (route simple tasks to cheap models, complex tasks to premium models) can now add a "high-complexity but cost-conscious" branch that points to grok-4.5 instead of defaulting to Opus-class or GPT 5.6 Sol.

2. Cached input economics. The $0.50/1M cached input rate is 25% of the standard input price. For coding agents that repeatedly load large context windows (repository trees, system prompts, long conversation histories), caching aggressively on Grok 4.5 is viable in a way that wasn't available at this performance tier. Teams using prompt caching on their gateway should evaluate Grok 4.5 as a primary path for long-context coding sessions.

3. Fallback design. Grok 4.5 is now available on the xAI API via standard OpenAI-compatible endpoints. Adding it as a fallback behind a Cursor-native or Anthropic-native primary is straightforward with any provider-agnostic routing layer. Given its strong performance on SWE Marathon — the benchmark that measures long-horizon, multi-step task success over time — it may outperform providers currently held as fallbacks in your configuration.

Provider dependency note. Grok 4.5 is a single-source model available only through the xAI API. Teams adding it to a routing stack should pair it with a fallback to a different provider family (Anthropic, Google, or a cloud-hosted alternative) for availability coverage.

What TheRouter users should watch or try

• Update your model allowlist: Add grok-4.5 to your provider configuration if you route to xAI. The model was confirmed live on the xAI API today.
• Run a token-count comparison: Pull a sample of your recent coding agent completions and estimate output tokens per resolved task. If your current provider averages well above 20K tokens on comparable tasks, Grok 4.5 is worth a cost-comparison run.
• Evaluate caching eligibility: If your coding agent reloads the same large system prompt or repository context repeatedly, the $0.50/1M cached input rate on Grok 4.5 could shift it to your preferred primary path for long-context sessions.
• Review fallback chains: The SWE Marathon result (29.0% pass@1) suggests Grok 4.5 holds up on longer autonomous runs — consider it as a primary or high-tier fallback rather than a secondary backup.

Grok 4.5 does not yet have EU availability (expected mid-July). Teams with regional routing requirements should hold off on making it a primary provider until regional endpoints are confirmed.

The token efficiency gap is the leading signal here. Most routing frameworks compare providers on benchmark score alone. The output token count per task is the number that actually appears on your invoice.