GPT-5.6 Terra vs Claude Sonnet 5: Benchmarks, Pricing, and Which to Use

If you're picking the mid-tier model to sit behind an agent that runs all day, the two obvious candidates right now are GPT-5.6 Terra and Claude Sonnet 5. Both are the default model their vendor hands to free users, both ship a roughly 1M token context window, and both land within a dollar of each other on input tokens.

In this article, I'll compare GPT-5.6 Terra and Claude Sonnet 5 across coding and terminal agents, knowledge work, cybersecurity posture, long-context behavior, API control surface, and pricing, so you can pick the one that fits your workload. For deeper coverage of each model on its own, see our guides to the GPT-5.6 family and Claude Sonnet 5.

TL;DR

• Terra is the stronger pick for terminal-driven coding agents and tool-heavy pipelines, where its token efficiency features do real work.
• Sonnet 5 is the calmer choice for long-context and document-heavy work, because its rates stay flat across the whole window.
• Choose GPT-5.6 Terra if your agents live in a shell, call many tools per turn, or you want cheaper output tokens after August.
• Choose Claude Sonnet 5 if you send very large prompts, want predictable per-token billing, or you deliberately want a model that declines exploit work.

What Is GPT-5.6 Terra?

GPT-5.6 Terra is the middle tier of OpenAI's GPT-5.6 family, which reached general availability on July 9, 2026, alongside the flagship Sol and the cheapest tier, Luna. OpenAI describes Terra as the balanced option for everyday work and notes that it roughly corresponds to the mini tier of earlier GPT-5 families. It ships a 1,050,000 token context window, a 128,000 token output ceiling, and a February 16, 2026, knowledge cutoff.

Two things make Terra interesting rather than just cheap. It inherits Programmatic Tool Calling from the Responses API, which lets the model write and run small in-memory programs that coordinate tools, rather than round-tripping each tool response through the model. And on July 30, 2026, OpenAI cut Terra's price by 20%, taking it from $2.50/$15 at launch to $2.00/$12 per 1M tokens.

We covered the wider family in our GPT-5.6 Sol, Terra, and Luna guide, and if you want a look at what the flagship tier is being used for, our write-up of how GPT-5.6 Sol helped disprove the Maxwell conjecture is a good read. There's also a flagship-level head-to-head in our Claude Opus 5 vs GPT-5.6 Sol comparison.

What Is Claude Sonnet 5?

Claude Sonnet 5 is Anthropic's mid-tier model, released on June 30, 2026, and pitched as the most agentic Sonnet the company has shipped. Anthropic's claim is that it plans, drives browsers and terminals, and runs autonomously at a level that previously needed Opus-class models, landing close to Opus 4.8 on reasoning, tool use, coding, and knowledge work at a lower price.

It is the default model on the Free and Pro plans, and it comes with a set of deliberate constraints: temperature, top_p, and top_k are gone, manual extended thinking returns a 400 error, and adaptive thinking is on by default. It also uses a new tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6, which Anthropic offset with introductory pricing through August 31, 2026.

We went through the launch numbers in our Claude Sonnet 5 guide and compared it upward against Anthropic's flagship in Claude Opus 5 vs Claude Sonnet 5.

GPT-5.6 Terra vs Claude Sonnet 5: Head-to-Head Comparison

Here is the shape of the two models side by side before we get into individual capability areas.

Coding and terminal agents

Usually, the split goes in Anthropic's favor for programming in a repository and in OpenAI's favor for terminal work.

The latter split remains: On Terminal-Bench 2.1, Terra reaches 87.4%, against 80.4% for Sonnet 5 in Anthropic's launch materials, a 7-point gap on command-line workflows. Notably, Terra also outperforms Claude Opus 4.8 (78.9%) and Claude Fable 5 (83.1%) in OpenAI's run of the same eval.

But currently, both competing mid-tier models are on par with one another for repo coding. On SWE-Bench Pro, which runs multi-step fixes against real repositories, OpenAI reports Terra at 63.4% and Anthropic reports Sonnet 5 at 63.2%.

Anthropic's own framing supports the split. Sonnet 5's coding story is about follow-through rather than peak score: early testers described it finishing tasks where Sonnet 4.6 stopped short and checking its own output unasked.

If your agent spends its life in a shell, Terra has the edge. If it spends its life inside a codebase reasoning about a failure, the benchmarks say you're picking between two models of the same class.

Knowledge work and professional tasks

Both vendors report GDPval-AA v2, an Elo-style measure of professional output quality, and the numbers are close enough to be a wash. OpenAI puts Terra at 1,593 Elo; Anthropic puts Sonnet 5 at 1618. OpenAI scores Claude Opus 4.8 at 1600 Elo on this eval while Anthropic scores the same model at 1615, so a 25-point gap between Terra and Sonnet 5 sits inside the methodology noise.

Where Terra has an unanswered claim is Agents' Last Exam, an evaluation of long-running professional workflows across 55 fields. Terra scores 50.4%, ahead of GPT-5.5 (46.9%), Claude Opus 4.8 (45.2%), and Claude Fable 5 (40.5%). Anthropic did not report Sonnet 5 on that eval, so treat it as OpenAI's chosen ground rather than a verdict.

Computer use is the one place I'd warn you off a direct comparison entirely. OpenAI reports Terra at 50.2% on OSWorld 2.0, while Anthropic reports Sonnet 5 at 81.2% on OSWorld-Verified. Those are different evaluation sets with different scoring, and Anthropic explicitly revised how it runs OSWorld-Verified for this release, so putting the two figures side by side tells you nothing.

Terra's practical advantage in this area is token accounting. Notion reported that many agents running GPT-5.5 perform just as well on Terra at half the cost and with 16% fewer tokens, and Clio reported that Programmatic Tool Calling cut prompt tokens by 38% on multi-step document analysis with no quality loss. Those are vendor-partner claims, not independent measurements, but they're specific about what changed.

Cybersecurity: two opposite bets

This is the widest gap in the comparison, and it is a design decision rather than a capability accident.

OpenAI trained the GPT-5.6 family for cyber work: Terra scores 91.8% on Capture-the-Flag challenges, 57.7% on SEC-Bench Pro, 52.9% on ExploitBench, and 23.2% on ExploitGym. Two of those figures beat GPT-5.5, the previous flagship, which scored 45.8% on SEC-Bench Pro and 47.9% on ExploitBench.

Anthropic went the other way. Sonnet 5 was not deliberately trained on cybersecurity tasks, and on an exploit-development eval built with Mozilla against Firefox 147 vulnerabilities, it never produced a working exploit at all, scoring 0.0%. Anthropic ships Sonnet 5 with real-time cyber safeguards enabled by default and recommends Opus 4.8 for cyber work that needs reduced guardrails.

For most readers, this is a feature of Sonnet 5, not a weakness. If you're building a product where a model attempting to generate exploits is a liability, a model that reliably declines is the safer dependency. If you do defensive security work, Terra sits inside OpenAI's Daybreak Trusted Access for Cyber program, and individual members need hardware-backed passkeys enabled by September 1 to keep access to the most cyber-capable models.

Worth flagging the friction cost on OpenAI's side, too. OpenAI says its GPT-5.6 Sol cyber safeguards block roughly 10 times as much potentially harmful activity as previous models, and it added a retry-on-a-lower-capability-model option in ChatGPT and Codex precisely because benign requests get caught. Reddit threads on the Anthropic side report the same complaint: Sonnet 5 is blocking benign security teaching content, so neither vendor has solved the overblocking issue.

Context window and long-context work

Nominally, these two are the same: Terra publishes a 1,050,000 token context window with a 128,000 token output ceiling, and Sonnet 5 ships a 1M window with a 128K output ceiling. What differs is what happens when you actually fill it.

Terra has published retrieval scores across the window. On OpenAI MRCR v2 8-needle, it scores 89.6% in the 256K-512K band and 72.5% in the 512K-1M band, close to Sol's 91.5% and 73.8% and far above Luna's 41.3%. This means that you can roughly expect the model to find hidden details in 9/10 cases within a 512k document, which makes it a very reliable companion for knowledge extraction. Anthropic did not publish comparable needle-retrieval figures for Sonnet 5 at launch.

The billing behavior is the more consequential difference. Prompts over 272K input tokens on Terra are priced at 2x input and 1.5x output for the entire request, not just the overage. Sonnet 5 has no published long-context surcharge, so its rates hold across the full window. I'll put numbers on that in the pricing section, because it flips which model is cheaper.

Control surface: effort, thinking, and API constraints

Both models expose a reasoning-effort dial, then diverge sharply on how much else you're allowed to touch.

Sonnet 5 removed temperature, top_p, and top_k entirely, and calls that set them return a 400 error. Manual extended thinking also returns a 400; you get adaptive thinking on by default, or you disable thinking outright.

Sonnet 5 also does not support Priority Tier on the Anthropic platform, which matters if your team relies on that for latency guarantees on Opus or Sonnet 4.6 today. Anthropic raised rate limits across Chat, Cowork, Claude Code, and the Claude Platform to absorb the heavier token usage that higher effort levels produce.

Terra's surface is wider in a different direction. The features worth knowing about:

• Programmatic Tool Calling in the Responses API, which runs coordinating programs in memory and is Zero Data Retention compatible.

• A multi-agent beta in the Responses API, letting the model run concurrent subagents and synthesize their work in one request.

• The max effort setting, toggleable by any user with GPT-5.6 access in ChatGPT Work and Codex.

• Explicit cache breakpoints with a 30-minute minimum cache life, cache reads at a 90% discount, and cache writes at 1.25x the uncached input rate.

• Function calling, structured outputs, hosted shell, code interpreter, and computer use on the Responses API, with fine-tuning not supported.

One asymmetry to note: OpenAI's ultra setting coordinates four agents in parallel, but every published ultra benchmark belongs to Sol, not Terra. Don't assume the mid-tier gets the same multi-agent gains until someone publishes Terra Ultra numbers.

Pricing: what you actually pay

Terra and Sonnet 5 currently sit within a dollar of each other on input, which makes the workload shape, not the headline rate, the thing that decides your bill.

Token rates side by side

Terra's $2.00/$12.00 rates reflect the 20% cut OpenAI applied on July 30, 2026, down from $2.50/$15.00 at launch. Cached input reads on Terra are $0.20 per 1M, and Anthropic supports prompt caching for Sonnet 5 without breaking out a Sonnet-specific rate in its launch materials.

Since Anthropic bills for thinking tokens at the output rate, running Sonnet 5 at extra-high effort is a direct multiplier on your most expensive line item. That is the single easiest way to blow through a budget you sized on input volume.

What a real workload costs

The formula is simple: (volume ÷ 1M) × rate, summed across input and output. I've worked three workload shapes at standard rates, chosen because each one isolates a mechanism where these two models actually differ.

The third row is the finding. Nothing changes about the total volume between rows two and three, only the size of each individual request, and Terra's 272K surcharge turns a 41% cost advantage into a 29% cost penalty. If you routinely stuff 300K tokens of contract text or repository context into a single call, Terra's cheaper sticker price is misleading.

The generation-heavy row also reverses if you're buying before September. At Sonnet 5's introductory $2.00/$10.00 rates, that workload costs $42 against Terra's $50, so Sonnet 5 is $8 cheaper until the standard rates kick in on September 1, 2026. Anthropic set that introductory window to make migration from Sonnet 4.6 roughly cost-neutral, which is worth remembering before you build a budget on it.

Why the same task costs different amounts

These rates only translate into a bill if both models consume the same number of tokens for the same work, and they usually don't. Terra and Sonnet 5 use different tokenizers from different vendors, so identical text yields different input token counts for each model before either model responds.

Neither vendor has published a between-subjects token count, so I can't hand you an adjusted figure. What exists are predecessor claims that don't say anything about Terra versus Sonnet 5:

• Anthropic documents Sonnet 5's new tokenizer producing roughly 30% more tokens than Sonnet 4.6, and
• Notion reported Terra using 16% fewer tokens than GPT-5.5.

My advice is to treat the table above as a rate comparison rather than a forecast, and to measure your own token counts on a representative sample before committing. The Sonnet 5 tokenizer change in particular caught teams out at launch, with Simon Willison's review and cloud consultancies both flagging it as the main migration surprise.

When to Choose GPT-5.6 Terra vs Claude Sonnet 5

Here's how the decision falls out across the scenarios I'd expect most readers to be in.

Choose GPT-5.6 Terra if...

• Your agents live in a shell. The 7-point Terminal-Bench 2.1 gap is the largest published coding difference between these two models.
• You call a lot of tools per turn. Programmatic Tool Calling keeps intermediate data out of the model context, and PlayCo reported 63.5% fewer total tokens and 50.1% fewer model turns on structured scene-construction work.
• You do defensive cyber work. Terra's 57.7% on SEC-Bench Pro and 52.9% on ExploitBench are in a different class from Sonnet 5's deliberate refusal.
• You need output tokens more cheaply. After August 2026, Sonnet 5's rate moves to $15 per 1M against Terra's $12.
• You want documented long-context retrieval numbers, like 72.5% on MRCR v2 in the 512K-1M band.

Choose Claude Sonnet 5 if...

• Your prompts or file context are large. Flat pricing to 1M tokens beats Terra's 272K cliff, and the difference is a 29% swing on a 10M input month.
• You want a model that declines cyber work by design. Anthropic's own assessment is that Sonnet 5 has substantially poorer cyber capabilities than Opus 4.8 and Mythos 5, which for most product teams is the desired behavior.
• You're already in the Claude ecosystem. Sonnet 5 is the default on Free and Pro and runs in Claude Code.
• You want a single dial rather than a settings panel. Adaptive thinking plus effort levels is a smaller surface to reason about than effort plus max plus ultra plus tool-calling modes.

If cost is the binding constraint, the three worked workloads are the fastest way to see which side you're on: $50 versus $63 on a generation-heavy month, $32 versus $45 on sub-threshold retrieval, and $58 versus $45 once each request crosses 272K input tokens.

How to Get Started With GPT-5.6 Terra and Claude Sonnet 5

Both models are generally available today, and both are the models their vendors give away for free, so trying them costs nothing but time.

Making your first API call

These are separate SDKs, so switching between them is more than a model string. Both calls are short enough that the difference is visible at a glance.

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-5.6-terra",  # swap for gpt-5.6-sol or gpt-5.6-luna
    input="Refactor this pytest fixture to avoid module-level state.",
)
print(response.output_text)

from anthropic import Anthropic

client = Anthropic()
message = client.messages.create(
    model="claude-sonnet-5",  # no temperature/top_p/top_k on this model
    max_tokens=1024,
    messages=[{"role": "user", "content": "Refactor this pytest fixture to avoid module-level state."}],
)
print(message.content.text)

If you're porting existing Anthropic code, remember to strip temperature, top_p, top_k, and any manual extended thinking configuration first, since Sonnet 5 returns a 400 rather than ignoring them.

For the full setup path on the Anthropic side, our complete guide to the Claude API goes deeper than makes sense here. To get started on the GPT side, our Working with the OpenAI API course is the best place to start.

Final Thoughts

The one-sentence version: Terra is the better mid-tier model for terminal agents and tool-dense pipelines, and Sonnet 5 is the better one for large prompts and teams that want a smaller surface to reason about.

What I find most interesting is that both vendors chose the mid-tier as the free default, and then made opposite bets about what a mid-tier model should refuse. OpenAI trained Terra for cyber work and put the sharpest capability behind identity verification and hardware passkeys. Anthropic trained Sonnet 5 not to do it at all, then shipped safeguards on top. Neither approach is obviously right, and both create real friction for legitimate work.

If you want to build the background to make these calls yourself rather than reading someone else's benchmark table, I recommend starting with our AI Fundamentals skill track.