Model comparison · July 10, 2026

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

OpenAI

GPT‑5.6 Terra

$2.50 / $15
Input / output per 1M tokens · 1.05M context · 128K max output
Anthropic

Claude Sonnet 5

$3 / $15
Standard input / output per 1M · 1M context · 128K max output
Short answer: choose Terra when terminal-agent coding and OpenAI’s Responses API / MCP workflow are central. Choose Sonnet 5 when Claude’s tool stack, adaptive-thinking workflow, and broadly available API access matter more—or while its introductory $2 / $10 rate lasts through August 31, 2026.

Price is close—but timing matters

API list pricingGPT‑5.6 TerraClaude 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 surchargeRequests over 272K: 2× input, 1.5× output for the request1M 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.

A simple 10M-input / 1M-output workload
Fresh tokens, no cache, no thinking-token estimate. This is an arithmetic illustration—not a task-cost benchmark.
Sonnet 5 intro
$30
Terra list
$40
Sonnet 5 standard
$45

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.

Coding benchmarks higher is better
Each value is the vendor-published score for that named benchmark. The chart deliberately does not combine them into a single “winner” score.
SWE‑bench Pro · Terra
63.4%
SWE‑bench Pro · Sonnet
63.2%
Terminal‑Bench · Terra
87.4%
Terminal‑Bench · Sonnet
80.4%
Terminal workSWE‑ProOutput priceContext/outputInput price
Radar: a visual index, not a benchmark

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

No published head-to-head MCP Atlas result: OpenAI has not published MCP Atlas results for Terra; Anthropic did not publish a Sonnet 5 MCP Atlas result. Both models support MCP-related workflows, but capability support is not a benchmark score. Likewise, do not infer a winner from a benchmark reported only by one vendor.

Product and workflow differences

Choose Terra if…

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.

Choose Sonnet 5 if…

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.

Run both if…

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