The one-line version: the cheapest and most expensive agent-capable models are now about 35x apart on output tokens, so the money question stopped being "which model?" and became "which model per workload?" Here is the August 2026 map — four tiers, real per-token prices — plus the one config change that lets you move between them when the prices move. And they move: OpenAI cut two GPT-5.6 tiers on July 30, and Claude Sonnet 5's introductory price expires August 31.
The map, at a glance#
Prices are per 1M tokens, input / output, as reported for early August 2026. Treat them as anchors, not gospel — pages move faster than blog posts.
| Tier | Models (in / out) | Best-fit workload |
|---|---|---|
| Budget / open-weight | DeepSeek V4 Flash 0731 ~$0.14 / $0.28 · GPT-5.6 Luna $0.20 / $1.20 | Bulk extraction, classification, tagging, background loops |
| Cheap-managed | Gemini 3.6 Flash $1.50 / $7.50 · Grok 4.5 $2 / $6 · GPT-5.6 Terra $2 / $12 | User-facing chat, single-step tool calls, RAG answers, routing |
| Default | Claude Sonnet 5 $2 / $10 → $3 / $15 Sep 1 · Kimi K3 $3 / $15 (cache hit $0.30) | Multi-step tool use, orchestration, code review |
| Frontier | GPT-5.6 Sol $5 / $30 · Claude Opus 5 $5 / $25 | Long-horizon coding, planning, high-stakes calls |
That is the whole answer for a founder skimming: pick the lowest tier that clears your quality bar for that specific job, not the one with the best reputation. The rest of this is the reasoning and the caveats.
Why "per workload" beats "per model"#
A year ago the sensible default was "use the best model you can afford and stop thinking about it." That logic broke when the budget tier started clearing real agent work. When a model at $0.28 per million output tokens handles your extraction job as well as one at $30, running that job on the frontier model isn't caution — it's a 100x overspend on the part of your pipeline that needed it least.
Most agent systems are a mix: a lot of cheap, high-volume calls (parse this, classify that, summarize the thread) and a few expensive, judgment-heavy ones (plan the refactor, decide the escalation). The cost win comes from not paying frontier prices for the cheap calls — and the quality win comes from not running the hard calls on a budget model to save pennies. Both mistakes are common; the fix is the same table above.
The cheapest model and the most expensive one are 35x apart on output. At that spread, "just use the best model" is a budgeting decision you're making by accident.
The calendar is the forcing function#
Two dated events make this month the time to re-price:
- August 31 — Claude Sonnet 5's introductory $2/$10 ends; from September 1 it's $3/$15, a 50% jump (Anthropic). If Sonnet 5 is your default backend, your bill rises next month whether or not you touch your code. The full breakdown is in Sonnet 5's intro pricing ends August 31 — what it does to your agent bill.
- July 30 — OpenAI cut GPT-5.6 Luna 80% to $0.20/$1.20 and Terra 20% to $2/$12 (CNBC). The budget and cheap-managed tiers just got cheaper, widening the gap you're leaving on the table by defaulting to premium.
The through-line is the same one this week's Wire called: the cheap tier is now the sensible default for agent work, and the premium tier is opt-in.
Three caveats that make these numbers lie#
- Price is not cost. An open-weight model you self-host (DeepSeek V4 Flash, Kimi K3) trades API dollars for inference operations and reliability engineering. Sometimes that's a win at volume; sometimes it's a second job. Count the ops.
- Cache changes the math. DeepSeek's first-party API reports a ~98% cache discount; Kimi K3's cache-hit input is $0.30 versus $3.00 cold. For repetitive agent prompts (the same system prompt on every call), the effective price can be a fraction of the sticker.
- Cross-vendor benchmarks aren't comparable. Two labs' numbers come from different harnesses, so a two-point gap is a tie. The only ranking that pays your bill is your own eval on your own tasks — task success rate, tool-call validity, and cost per completed task, not cost per token.
The one change that makes the map usable#
None of this matters if switching models means a code change. Put a swappable client in front of every LLM call: one interface, the model chosen by workload and config, never hardcoded at the call site. Then routing bulk work to the budget tier and reserving the frontier for the calls that earn it is a config edit, not a refactor — and when the next price cut lands (two arrived in the last week), you move volume without shipping code.
If you want the deeper decision logic behind the table: when it still pays to run the flagship, the Grok 4.5 vs GPT-5.6 vs Opus 4.8 coding head-to-head, and Kimi K3 vs Opus vs GPT-5.6 on cost. But the move for this afternoon is smaller than any of them: price your workloads against the table, put a swappable client in front of them, and do it before August 31 — because on September 1 the default tier costs 50% more and the map redraws itself.
