The decision in one line: For the workload that eats most of your API bill — the model you call in a loop, not the one you demo — Gemini 3.5 Flash is the cheap default, Claude Sonnet 5 is the agent workhorse, and GPT-5.6 Terra is the coder. The rest of this piece is how to tell which one your workload actually wants.
Why the mid-tier is the only tier that matters to your budget#
Founders obsess over frontier models and pay for mid-tier ones. The frontier model writes your launch tweet; the mid-tier model runs ten thousand extraction calls, answers every RAG query, and drives every agent step. That's where the invoice comes from. So when all three labs refreshed this tier in one week, it mattered more to a small team's runway than any benchmark headline.
Here's the sticker price, cleaned up:
| Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|
| Gemini 3.5 Flash | $1.50 | $9.00 | 1M |
| Claude Sonnet 5 (intro) | $2.00 | $10.00 | 200K |
| GPT-5.6 Terra | $2.50 | $15.00 | 400K |
Sonnet 5's intro rate holds through August 31, then rises to $3/$15 — the same output price as Terra. Plan for the standard number if your migration outlasts the summer.
The cost mistake everyone makes: reading input price#
Input price is the number that catches your eye and the wrong one to optimize. Most production LLM work — agents especially — is output-bound: a short instruction produces a long tool call, a multi-paragraph answer, a chain of reasoning. When output dwarfs input, the output rate is your bill.
On output price the order is clean: Flash ($9) < Sonnet 5 intro ($10) < Terra ($15). Terra costs two-thirds more per output token than Flash. Over a million agent steps that is not a rounding error; it's a hiring decision.
There is one common exception. If you're doing long-context retrieval — stuffing 200K tokens of documents in to get a short answer out — the workload flips to input-bound, and Flash's $1.50 input plus a 1M window and $0.15 cached-input rate makes it the obvious pick. Match the price you optimize to the shape of your traffic.
The Sonnet 5 asterisk: cheaper per token, maybe not per task#
The one trap in this table is invisible in it. Sonnet 5 ships a new tokenizer that can map the same text to roughly 1.0–1.35x more tokens than before.
Per-token price only translates to cost-per-task if the token count is stable. If your prompts hit the top of that range, Sonnet 5's effective cost drifts up toward — or past — its rate-card advantage over Flash. This isn't a reason to avoid it; near-Opus quality at this price is a real offer. It's a reason to never migrate on the rate card alone. Replay a representative sample through both models, count the tokens each actually bills, and compare total spend on your traffic. The rate card is marketing; your trace is the truth — and it takes about forty lines of code to capture that trace.
Pick by the job, not the benchmark#
Reach for Gemini 3.5 Flash when:
- Volume is high and the task is bounded — classification, extraction, tagging, routing.
- You're doing long-context RAG; the 1M window and cheap cached input are decisive.
- You want the lowest floor on cost and can tolerate a slightly weaker reasoning ceiling.
Reach for Claude Sonnet 5 when:
- You're running agents — tool use, multi-step loops — where its near-Opus behavior earns back the price versus a weaker model that fumbles the trajectory.
- Quality-per-dollar on reasoning matters more than the absolute cheapest token.
- (Just validate the token math first.)
Reach for GPT-5.6 Terra when:
- Code generation or code-adjacent reasoning is the core task; it's the strongest coder of the three and slots straight into Codex.
- You want one everyday model that's "5.5-level at half the cost" and you're already in the OpenAI ecosystem.
- You need a 400K context but not the full million.
The move that outlasts this comparison#
Every number here has a shelf life measured in weeks. Sonnet 5's intro pricing expires August 31. Gemini 3.5 Pro is still landing. GPT-5.7 is a matter of time. The teams that win this churn aren't the ones who pick perfectly today — they're the ones who put a thin routing layer between their app and the model so switching is a config change, not a refactor. Route cheap tasks to the cheap model, hard tasks to the strong one, and keep every provider one line away. Then let the labs fight the price war for you.
