Claude Sonnet 5 vs GPT-5.6 Terra: The Workhorse-Tier Verdict
Claude Sonnet 5 ($2/$10 intro through August 31) and GPT-5.6 Terra ($2.50/$15) are fighting for the mid-tier where most production LLM spend actually lands. Sonnet 5 posts an 85.2% SWE-bench Verified score while OpenAI no longer reports Terra to that board, and Anthropic exempts cached reads from input rate limits. The catch: Sonnet's standard $3/$15 pricing arrives September 1, handing the input-cost edge to Terra. For PMs, the choice comes down to workload shape — cache-heavy agentic work favors Sonnet 5, input-heavy stable budgets favor Terra.
Your model line item is probably a workhorse-tier line item, and both vendors just repriced it. This week, run a two-column cost projection for your highest-volume workload: one column at Sonnet 5's intro rate of $2/$10 per million tokens, one at the $3/$15 standard rate that lands September 1, and set both against Terra's flat $2.50/$15, as of July 2026. The exercise takes an hour with your token logs and settles the argument with data instead of vendor loyalty. Look past list price while you're in there. If your architecture reuses large prompts — agent scaffolds, long system contexts, RAG headers — Anthropic's exemption of cached reads from input rate limits is worth real throughput, per its docs, and Terra has no documented equivalent. If your workload is input-heavy and cache-light, Terra's $2.50 input beats Sonnet's post-August $3 and never expires. To be fair to OpenAI, an unpublished benchmark isn't a failed one; Terra may hold up fine on your own evals, and its price needs no calendar footnote. But you can't verify that from the outside, which is itself information when you're defending a platform choice. The intro window closes August 31. Every week you defer this analysis, you burn margin you could have banked and shrink your runway to migrate before the math changes underneath you.
Claude Sonnet 5 costs $2/$10 per million tokens through August 31 versus Terra's $2.50/$15, then moves to $3/$15 — handing the input-price advantage to OpenAI on September 1.
Twenty percent. That's the input-price gap between Claude Sonnet 5 ($2 per million input tokens) and GPT-5.6 Terra ($2.50) as of July 2026 — and on September 1, it inverts. My verdict up front: Sonnet 5 is the stronger workhorse pick today, on published evidence, caching mechanics, and raw price.
Terra is the safer choice for teams that budget on stability instead of promotions. The flagship fight gets the headlines, but this tier is where most production token spend actually lands, and both companies are pricing like they know it.
The pricing math has an expiration date
Anthropic shipped Sonnet 5 on June 30 at $2 per million input tokens and $10 per million output — an introductory rate that holds through August 31, 2026, before standard pricing of $3/$15 takes effect, per Anthropic's pricing docs. OpenAI took GPT-5.6 Terra to general availability on July 9 at $2.50/$15, per OpenAI's pricing page, with no deadline attached.
Run the arithmetic both ways. Today, Sonnet 5 undercuts Terra by 20% on input and 33% on output. On September 1, Sonnet's input price climbs to $3, which is 20% above Terra's, while output equalizes at $15.
A workload that skews input-heavy, like long-document analysis or retrieval pipelines, flips from favoring Anthropic to favoring OpenAI in one billing cycle. I walked through the deadline mechanics in our pricing-deadline explainer; the short version is that any cost model you build this month needs a second column dated September 1.
The asymmetry is deliberate. The $2/$10 window is widely read as IPO-timed aggression, a case our launch analysis makes in detail. Anthropic is buying workhorse-tier share at a discount; OpenAI is betting that predictability wins procurement.
One model publishes benchmarks. The other stopped.
Sonnet 5 posts 85.2% on SWE-bench Verified per the llm-stats board (July 14, 2026, self-reported scores) and 63.2% on SWE-bench Pro per launch-week reporting. Terra posts nothing on that board. OpenAI stopped self-reporting to it, and the company's own framing — that Terra performs competitively with its pricier predecessors, per OpenAI — is a relative claim, not a score.
To be fair to them, OpenAI's retreat from public leaderboards isn't baseless. Self-reported benchmark tables carry real contamination and cherry-picking problems, and opting out is arguably more honest than quietly optimizing for the test set. But for a builder allocating budget, comparing these two models is like judging a race where one runner posts an official time and the other assures you he was fast.
Sonnet 5's 85.2% is self-reported, yet it is public and falsifiable. Terra's coding ability, as of July 2026, lives in private evals and anecdotes.
If your team runs its own eval harness, this gap costs you an afternoon. If you need published numbers to justify a platform bet to finance, it costs you the argument.
Context, caching, and the rate-limit fine print
| Spec (as of July 2026) | Claude Sonnet 5 | GPT-5.6 Terra |
|---|---|---|
| Price per 1M tokens (in / out) | $2 / $10 intro; $3 / $15 from Sep 1 | $2.50 / $15 |
| Context window | 1M tokens | 1.05M tokens |
| Max output | 128K (300K via Batches beta) | 128K |
| SWE-bench Verified | 85.2% (llm-stats) | Not reported |
| Cache reads | Up to 90% off base input | Discounted (per OpenAI docs) |
| Cached reads vs rate limits | Exempt from input-token limits | No documented exemption |
Terra's 1.05M context edges Sonnet's 1M by 50K tokens, a margin most workloads will never touch. On the Anthropic side, cache reads run up to 90% below base input and cache writes bill at 1.25x, per its pricing docs; OpenAI discounts cached input on Terra as well. The real separation is in the plumbing.
On the Claude API, cached reads don't count toward input-token rate limits at all — a documented Anthropic advantage that materially raises effective throughput for cache-heavy agent loops, because the repeated scaffold of every loop iteration stops consuming quota.
Sonnet 5 also gets its own dedicated rate bucket: 1,000 requests and 2M input tokens per minute at the Start tier, 5,000 and 5M at Build, per Anthropic's rate-limit tables, separate from the combined Sonnet 4.x pool. Teams coming off Sonnet 4.6 get fresh headroom rather than shared quota; our migration guide covers the rest of that move. These numbers shift often, so check the live model tracker for current figures on both models.
Who should pick which
Pick Sonnet 5 if you're building agentic coding systems, your traffic is cache-read-heavy, or you need a public benchmark record to defend the decision internally. Through August 31 it is cheaper on every axis, and even at $3/$15 the cache-read exemption from rate limits can matter more than list price for loop-heavy agents.
Pick Terra if your spend skews input-heavy and must stay predictable past September 1, if you're already deep in OpenAI's tooling, or if you won't run procurement math against a promotional clock.
My read on what happens next: OpenAI responds before the window closes, with either a Terra price cut or a cache-exemption match by September 1. Which leaves builders with the question no pricing page answers: when the workhorse tier turns into a promotional battleground, do you architect for the cheapest model this quarter, or for the vendor whose pricing behavior you can still predict next year?
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
Treat it as strong but not neutral evidence. The score sits on the llm-stats board (as of July 2026) with self-reported methodology, which means it's public, comparable to prior Claude models, and falsifiable by anyone who reruns the harness. That's categorically more useful than Terra's situation, where OpenAI stopped reporting to the board entirely and offers only relative claims against its own earlier models. The practical move is to run both models through your own eval set on your actual tickets before committing production traffic.
It depends on the date and the shape of your traffic. Through August 31, 2026, Sonnet 5 wins outright at $2/$10 versus Terra's $2.50/$15, and Anthropic's Batch API takes a further 50% off both input and output. From September 1, Sonnet's $3/$15 standard rate means Terra is 17% cheaper on input while output ties at $15, so input-heavy pipelines tilt toward OpenAI. Cache-heavy agent workloads are the exception: Anthropic discounts cache reads up to 90%, and it alone exempts cached reads from input rate limits, which raises effective throughput without raising spend.
The pricing reset. Any unit economics built on the $2/$10 intro rate overstate margin by up to 50% on both token types once standard pricing lands September 1, 2026. There's also concentration risk: features built around Sonnet-specific advantages like the cache-read rate-limit exemption or the 300K batch output beta get harder to port later. Mitigate both by keeping a provider abstraction layer in place and re-approving the decision against September numbers, not July numbers.
