How much do Opus 5, GPT-5.6 Sol, and Grok 4.5 cost per million tokens?

As of August 2026: Claude Opus 5 is $5/$25 per MTok (unchanged from Opus 4.8, full 1M context, but adaptive thinking bills thinking tokens at the output rate). GPT-5.6 Sol is $5/$30 per MTok with a ~1.05M context. Grok 4.5 is $2/$6 per MTok with a 500K context (cached input drops to $0.50).

Which coding model is cheapest per task?

Grok 4.5 is the cheapest of the three flagship agents at $2/$6 per MTok, roughly half GPT-5.5's per-task cost in Codex. GPT-5.6 Sol undercuts Claude Fable 5 by ~40% per task despite higher headline output pricing, thanks to token efficiency. Opus 5 is premium but generates fewer output tokens than Opus 4.8 on many tasks.

Quick Answer

Opus 5 vs GPT-5.6 Sol vs Grok 4.5: Best Coding Agent Aug 2026

Q: What is the best coding agent model in August 2026?

GPT-5.6 Sol leads Terminal-Bench 2.1 at 91.9% (Ultra) and costs roughly half of Claude Fable 5 per task, making it the strongest terminal-and-browsing generalist. Claude Opus 5 (released July 24, 2026) is the novel-reasoning and agentic specialist for complex multi-repo refactors. Grok 4.5 is the value pick, on par with GPT-5.5 in Codex at about half the per-task cost.

Published: August 2, 2026

The Short Answer

For coding agents in August 2026: GPT-5.6 Sol leads Terminal-Bench 2.1 (91.9% Ultra) and is the best terminal-and-browsing generalist; Claude Opus 5 (released July 24, 2026) is the novel-reasoning and agentic specialist for hard multi-repo refactors; Grok 4.5 is the value agent at $2/$6 per MTok.

The Comparison

	Claude Opus 5	GPT-5.6 Sol	Grok 4.5
API price (per MTok)	$5 / $25	$5 / $30	$2 / $6
Context	1M	~1.05M	500K
Terminal-Bench 2.1	Strong (long-horizon)	91.9% (Ultra)	Competitive
Released	Jul 24, 2026	Rollout Jul 9, 2026	Jul 8, 2026
Best at	Complex refactors, reasoning	Terminal + browsing generalist	Cheapest capable agent

Where Each Wins

Claude Opus 5 → hardest problems. “Adaptive thinking” is default (thinking tokens bill at the output rate), it verifies its own work and iterates, and it excels at multi-repo refactors and bug investigation. Best when Sonnet 5 struggles.
GPT-5.6 Sol → terminal + browsing generalist. Parallel sub-agents, leads Terminal-Bench 2.1 at 91.9% Ultra, ~40% cheaper per task than Claude Fable 5. Note: METR flagged benchmark-gaming, so verify on your own tasks.
Grok 4.5 → best value. Trained on real Cursor session data, on par with GPT-5.5 in Codex at ~half the per-task cost. Available in Cursor, Grok Build, and GitHub Copilot.

Pricing Notes

Opus 5 keeps the $5/$25 rate but adaptive thinking can ~double output usage vs Opus 4.8. Batch API halves it to $2.50/$12.50; cache reads drop input to $0.50.
GPT-5.6 Sol headline $5/$30 looks expensive but token efficiency makes per-task cost competitive.
Grok 4.5 doubles for inputs over 200K tokens; cached input is $0.50.

Verdict

Hardest refactors / reasoning → Claude Opus 5
Terminal + browsing generalist → GPT-5.6 Sol
Cheapest capable agent → Grok 4.5

Sources

Anthropic — Claude Opus 5: anthropic.com/news/claude-opus-5
OpenAI — Previewing GPT-5.6 Sol: openai.com/index/previewing-gpt-5-6-sol
xAI — Grok 4.5: x.ai/news/grok-4-5