VynarisEarly betaGet your API key

Blog · 2026-07-16 · Vynaris Team

Claude Sonnet 5 stays at $2/$10: Anthropic cancels 50% rise

Claude Sonnet 5 stays at $2/$10 after Anthropic canceled its scheduled 50% increase. The $36 balanced-task bill will not become $54.

Claude Sonnet 5 will stay at $2 per million input tokens and $10 per million output tokens. Anthropic canceled the scheduled 2026-09-01 move to $3/$15. A balanced 1,000-task bill therefore stays $36 instead of reaching $54. Prices verified 2026-08-14.

_Updated 2026-08-14: this article originally analyzed the announced September increase. Anthropic changed the live pricing page, so we reversed the forecast and recomputed every table below._

TL;DR

Verdict table

Workload per 1,000 tasks                   Permanent $2/$10 bill  Canceled $3/$15 forecast  Increase that will not arrive
-----------------------------------------  ---------------------  ------------------------  -----------------------------
Balanced: 8M input + 2M output             $36.00                 $54.00                    $18.00
Balanced, Batch API                        $18.00                 $27.00                    $9.00
Cache-heavy: 2M miss + 6M hit + 2M output  $25.20                 $37.80                    $12.60
Agent step: 50M input + 15M output         $250.00                $375.00                   $125.00

The Anthropic pricing documentation now says the $2/$10 launch price “is now the standard price” and that the scheduled $3/$15 increase “will not occur.” That first-party correction replaces every earlier forecast in this article.

The $18 calculation

Our balanced task contains 8,000 input tokens and 2,000 output tokens. Across 1,000 tasks, that becomes 8 million input and 2 million output tokens.

At the permanent price, 8 × $2 + 2 × $10 = $36. The canceled price would have been 8 × $3 + 2 × $15 = $54. The difference is $54 - $36 = $18, or 33.3% below the old forecast.

The calculation is deliberately boring. It needs no traffic estimate, benchmark score or internal data. Change the token shape in the LLM cost calculator if your workload is output-heavy or carries a long context.

The Batch API remains 50% below standard rates. That makes the balanced bill $18, not the $27 we had projected for September. Batch is still suitable only when the workload can accept its asynchronous delivery window.

Cache and agent workloads

Prompt caching also keeps the existing base. For 1,000 tasks with 6 million cache-read tokens, 2 million uncached input tokens and 2 million output tokens, the bill is 6 × $0.20 + 2 × $2 + 2 × $10 = $25.20. The canceled schedule would have charged $37.80.

An agent step with 50,000 input and 15,000 output tokens costs $0.05 × $2 + $0.015 × $10 = $0.25. One thousand steps cost $250. The old forecast was $375. Teams that budgeted the announced rate can remove $125 per 1,000 such steps from the September reserve.

These are per-request cost calculations. They do not claim a task uses any universal token count. Measure your own requests before turning the deltas into a budget.

What the cancellation changes

The immediate answer is simple: no September price-triggered migration is required. A route that cleared its quality and cost gates at $2/$10 keeps the same sticker.

The cancellation also changes two prior comparisons. Our Claude Max versus API break-even now keeps Sonnet 5 at 592 one-hour sessions per $100 on its stated cached workload. The previously projected 395-session threshold is gone.

Our Anthropic tokenizer cost analysis still matters. Anthropic says Sonnet 5's newer tokenizer can produce about 30% more tokens for the same text than older models. The permanent lower sticker can offset that inflation. Token counts still decide the real bill.

Do not confuse a canceled increase with a guaranteed future price. “Standard” describes the live rate, not a lifetime contract. Keep a dated price snapshot beside every model routing policy and re-run the unit math when a provider edits its page.

Honest tradeoff

Staying on Sonnet 5 avoids migration work, but price stability does not prove it is the cheapest correct model. A cheaper route can still win on high-volume extraction or classification. Conversely, moving solely because of a canceled forecast now creates engineering risk without the expected saving. Re-evaluate on measured quality, not on yesterday's announcement.

FAQ

Does Claude Sonnet 5 still get more expensive on 2026-09-01?

No. Anthropic says the $2/$10 price is now standard and the scheduled $3/$15 increase will not occur.

What will 1,000 balanced Sonnet 5 tasks cost?

$36 for 8 million input and 2 million output tokens. Batch costs $18 on the same token totals.

Do prompt-cache rates also stay at the lower base?

Yes. A cache read remains $0.20 per million tokens on the live pricing table. The article's cache-heavy example stays at $25.20 per 1,000 tasks.

Vynaris option

Vynaris is an OpenAI-compatible inference gateway for teams that want pricing changes outside application code. One base URL swap lets a routing policy absorb a provider re-rate without rewriting each call site.

Source