GPT‑5.6 Terra vs Claude Sonnet 5
At standard API rates, these are almost a dead heat: Terra costs $2.50 / $15 and Sonnet 5 $3 / $15 per million input/output tokens. But the closer you look, the less interchangeable they become.
Last updated July 10, 2026 · Primary sources: OpenAI and Anthropic · Benchmark caveats included
GPT‑5.6 Terra
Claude Sonnet 5
Price is close—but timing matters
| API list pricing | GPT‑5.6 Terra | Claude Sonnet 5 |
|---|---|---|
| Input / 1M tokens | $2.50 | $3.00 standard $2.00 introductory through Aug. 31, 2026 |
| Output / 1M tokens | $15.00 | $15.00 standard $10.00 introductory through Aug. 31, 2026 |
| Cached-input read / 1M | $0.25 | $0.30 standard ($0.20 introductory) |
| Cache write / 1M | $3.125 (1.25× input) | $3.75 standard ($2.50 introductory) |
| Long-context surcharge | Requests over 272K: 2× input, 1.5× output for the request | 1M context at standard rates |
Sonnet 5 uses a new tokenizer that Anthropic says produces roughly 30% more tokens for the same text on average. Per-token price alone is therefore not a reliable per-document cost estimate—measure your own prompts.
Published coding results
Both vendors publish scores on SWE‑bench Pro and Terminal‑Bench 2.1. Those numbers are useful directional signals, but their agent configurations and evaluation scaffolds are vendor-reported; treat cross-lab comparisons as indicative rather than a precise rank order.
The radar is intentionally limited to published, transparent dimensions. Terminal and SWE‑Pro use the published scores above. “Input price” uses standard list price; context and output capacity are effectively tied. It is a comparison aid, not a scientific composite metric.
GPT‑5.6 Terra Claude Sonnet 5
- Terra’s visible edge: higher published Terminal‑Bench 2.1 score and lower standard input rate.
- Near-tie: SWE‑bench Pro differs by 0.2 percentage points—well below what should drive a purchasing decision without a shared harness.
- Capacity: both offer about one million tokens of context and 128K output tokens.
What cannot be honestly compared yet
Product and workflow differences
You use the OpenAI API, Codex, Responses API, hosted shell, or programmatic tool calling; you want the stronger published terminal-agent score; and the 30-minute minimum cache life / explicit cache breakpoints fit your traffic pattern.
You want a generally available Claude API model today, Claude’s adaptive-thinking and tool workflow, or its introductory pricing. Sonnet 5’s 1M context is standard-priced and it is available across Claude API, Bedrock, Google Cloud, and Microsoft Foundry.
You are choosing for production coding. Use a held-out set of your real repository tasks, track merged-patch acceptance, tool failures, latency, and billed tokens—not just a public benchmark.
Verdict
At standard pricing, Terra is the cleaner price/performance bet for OpenAI-native terminal agents: it is $0.50 cheaper per 1M input tokens, ties on output price, and has the higher published Terminal‑Bench 2.1 result. Sonnet 5 is not simply “the same price,” though: during its introductory window it is materially cheaper, and its tokenizer can change real cost calculations. The SWE‑bench Pro gap is effectively a tie on vendor-reported data. Select on tooling, availability, and your own eval—not a 0.2-point benchmark delta.
Sources & methodology
- OpenAI — GPT‑5.6 launch, specs, pricing, and evaluation table (Terra pricing; 1.05M context; published coding scores).
- OpenAI API model documentation — GPT‑5.6 Terra (128K maximum output and cache pricing).
- Anthropic — Introducing Claude Sonnet 5 (availability and introductory pricing).
- Anthropic docs — What’s new in Sonnet 5 (standard pricing, tokenizer, 1M context, 128K output).
- Claude Sonnet 5 System Card (published SWE‑bench Pro and Terminal‑Bench 2.1 scores).
- Scale MCP Atlas leaderboard (used only to check whether a result exists; none was found for either model).