LumeAPILast verified: August 15, 2026
Short path: AI API pricing · Claude Sonnet 4.6 · Claude Opus 4.8 · Cheap LLM pick list
LumeAPI is an independent third-party gateway — not Anthropic. This page owns Sonnet vs Opus on the same loop. Agent hop caps live on the agent bill guide. The three-provider matrix lives on the 2026 pricing comparison.
Quick Answer
| Decision | Catalog id on LumeAPI | Anthropic standard $/1M in/out | LumeAPI $/1M in/out |
|---|---|---|---|
| Default production volume | claude-sonnet-4-6 | $3.00 / $15.00 | $1.50 / $7.50 |
| Escalate after a logged eval fail | claude-opus-4-8 or claude-opus-5 | $5.00 / $25.00 | $2.50 / $12.50 |
| Need Sonnet 5 at $2 / $10 | not in LumeAPI catalog | $2.00 / $10.00 (now standard) | — use Anthropic directly |
- Run the task class on Sonnet 4.6. Log accept/fail.
- Send only the failed class to Opus. Do not default 20% of traffic to Opus.
- If prompt cache or Batch is the real saving, stay on Anthropic — those are not LumeAPI Chat Completions discounts.
In short
Opus is a premium hop, not a default. LumeAPI lists matching Sonnet 4.6 and Opus 4.8/5 standard text rates at 50% of Anthropic's base input/output. Sonnet 5 ($2 / $10 on Anthropic as of this verification) is a different model and is not in the LumeAPI catalog — do not treat it as a 50% off Sonnet 4.6.
What most guides get wrong
They mix Sonnet 5 with Sonnet 4.6, and Batch/cache with real-time. Anthropic's pricing page now states the Sonnet 5 $2 / $10 rate is standard (the planned September 1, 2026 rise to $3 / $15 will not occur). LumeAPI still lists claude-sonnet-4-6 at $1.50 / $7.50. That is not “Sonnet 5 at half price.”
Claude 4.7 and later also use a newer tokenizer that Anthropic says produces about 30% more tokens for the same text. Token-count comparisons across generations must use billed usage, not word count.
Same loop, two catalog ids
Illustrative agent task: 40,000 input + 5,000 output tokens, one attempt, no cache, no Batch.
LumeAPI Sonnet 4.6:
40,000 / 1e6 × $1.50 + 5,000 / 1e6 × $7.50 = $0.0600 + $0.0375 = $0.0975LumeAPI Opus 4.8 (or Opus 5 — same catalog rates):
40,000 / 1e6 × $2.50 + 5,000 / 1e6 × $12.50 = $0.1000 + $0.0625 = $0.1625Opus premium on this shape: $0.065 per attempt. Opus wins only if avoided retries, review minutes, or failed-task recovery exceed $0.065. Hypothetical arithmetic; not a customer result.
| Cost to recover one Sonnet failure | Required extra Opus acceptance (this token shape) |
|---|---|
| $0.25 | 26 percentage points |
| $1.00 | 6.5 percentage points |
| $5.00 | 1.3 percentage points |
If Sonnet already accepts 90% and recovery is $1, Opus must clear 96.5% on the same eval or it does not pay for itself on tokens alone.
When Anthropic-native cache beats the gateway
Anthropic documents cache hits on Sonnet 4.6 at $0.30 / 1M input vs $3.00 base (platform pricing). LumeAPI Chat Completions debit $1.50 / 1M input for claude-sonnet-4-6 with no Anthropic prompt-cache product.
If a stable prefix is cache-hit on most calls, the official API can beat the gateway on input even before Batch. Measure cache-hit rate on Anthropic before you move that workload. Batch is a separate 50% product for async jobs — do not apply it to interactive agents.
Prove the Sonnet id
curl https://api.lumeapi.site/v1/chat/completions \
-H "Authorization: Bearer $LUMEAPI_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"claude-sonnet-4-6","messages":[{"role":"user","content":"ping"}],"max_tokens":16}'The Anthropic Messages API is a different shape. LumeAPI's documented path is OpenAI-compatible Chat Completions. An Anthropic SDK app is not a three-field swap.
A realistic production scenario
A coding queue: 10,000 tasks/month, 8,000 pass Sonnet on the first try, 2,000 fail a unit-test rubric.
- 8,000 × $0.0975 ≈ $780 on Sonnet (illustrative token shape above)
- Escalate only the 2,000 failures to Opus at $0.1625 ≈ $325
- Default-everything-Opus would be 10,000 × $0.1625 = $1,625
Selective escalation is $1,105 vs $1,625 in this worksheet. If Opus does not raise accept rate on those 2,000, you paid the premium for nothing — keep the fail class on a repaired Sonnet prompt instead.
Expert take
Default claude-sonnet-4-6. Log the fail class. Queue Opus only there. If you need Sonnet 5, cache, or Batch, that decision is Anthropic-native — not a LumeAPI catalog row.
FAQ
What is Sonnet vs Opus pricing on LumeAPI?
Sonnet 4.6 is $1.50 / $7.50 per 1M in/out. Opus 4.8 and Opus 5 are $2.50 / $12.50. Spec pages: /models/claude-sonnet-4-6, /models/claude-opus-4-8.
Is Claude Opus 5 too expensive for production volume?
As a default, yes for most queues. Use the $0.065 premium test above. Escalate when recovery cost × extra accept rate exceeds that premium.
Should I use Sonnet 5 or Sonnet 4.6?
Sonnet 5 is $2 / $10 on Anthropic and is not in the LumeAPI catalog. Sonnet 4.6 on LumeAPI is $1.50 / $7.50. Pick by eval, not by name similarity.
Is LumeAPI always 50% cheaper than Anthropic?
No. The 50% figure is same-model standard real-time text rates in the table. Cache hits, Batch, Sonnet 5, and contracts can beat it.
Can I keep the Anthropic SDK?
Not as a documented LumeAPI path. Change to the OpenAI SDK, base_url=https://api.lumeapi.site/v1, and a catalog Claude id.
Sources and methodology
- Anthropic platform pricing — verified August 15, 2026: Opus 5/4.8 $5 / $25; Sonnet 4.6 $3 / $15; Sonnet 5 $2 / $10 standard; cache-hit column for Sonnet 4.6 $0.30 / 1M.
- LumeAPI pricing catalog — catalog sync 2026-07-22.
- LumeAPI models
No private benchmark is claimed. For the live two-column hub see /ai-api-pricing.