Did Claude Sonnet 5 pricing go up on September 1, 2026?

No. Anthropic cancelled the increase on August 10, 2026, three weeks before it was scheduled to take effect. Sonnet 5 launched June 30 at $2 per million input tokens and $10 per million output, explicitly as introductory pricing through August 31, with standard pricing of $3/$15 set to follow. Anthropic's Claude Platform Docs now state that the $2/$10 rate is permanent and the scheduled September 1 increase to $3/$15 will not occur.

Why did Anthropic cancel the Claude Sonnet 5 price increase?

Anthropic hasn't published a reason, but the timing lines up with broader mid-tier price competition. Three weeks before the cancellation, on July 30, 2026, OpenAI cut GPT-5.6 Luna's price 80% and Terra's 20%, just three weeks after that model family launched. The frontier LLM token price index sat at roughly 12 in August 2026, 88% below its March 2023 baseline, with mid-tier models under the most downward price pressure. Raising Sonnet 5 by 50% into that environment would have pushed it out of the price band it competes in.

What should teams do if they already changed their routing to avoid the September price hike?

Reverse it. If a team pre-emptively shifted default traffic away from Claude Sonnet 5, expecting a 50% price jump on September 1, that traffic is now paying a different model's price and possibly a quality cost for an increase that never happened. The corrective action is the same as the original mistake in reverse: check the current, real rate before making the routing decision, rather than acting on a rate that was only ever announced, not yet charged.

Should LLM pricing be hardcoded into application code or routing config?

No. Hardcoded per-token price constants go stale the moment a provider changes its mind, and 2026 has shown that providers change pricing on a shorter, less predictable cycle than they change models. Anthropic both discounted and then permanently kept Sonnet 5's launch price within the same two-month window, and separately closed its Priority Tier to new commitments mid-year. A price should be read from a live, versioned source at request or dashboard-render time, the same way a routing decision reads current model availability rather than a value fixed when the integration was written.

How does a cancelled price increase affect LLM cost dashboards and forecasts?

Any dashboard or forecast that hardcoded the announced $3/$15 rate for Claude Sonnet 5 starting September 1 is now overstating future cost, potentially by 50% on every Sonnet 5 request in the forecast. A system that prices requests against live provider rates instead of a cached constant reflects a reversal like Anthropic's automatically, the same day it takes effect, without anyone needing to manually update a spreadsheet or re-check the provider's pricing page.

The Cliff That Didn't Happen

On June 30, 2026, Anthropic launched Claude Sonnet 5 at an introductory $2/$10 per million tokens, with a $3/$15 hike due September 1 that never came. We wrote about that countdown in July. Anthropic's pricing docs now say the $2/$10 rate is permanent and the September hike "will not occur," reversed on August 10, three weeks before it would have taken effect and right in the middle of OpenAI cutting GPT-5.6 Luna's price 80% on July 30. Nothing about the model changed twice in six weeks; only the number on a pricing page did, in both directions. This post covers what happened, why a mid-tier price hike got cancelled in a market where frontier prices sit 88% below their 2023 baseline, and why a hardcoded price constant is now a bigger liability than most teams' hardcoded model names.

On August 10, 2026, Anthropic cancelled a price increase it had told everyone was coming.

Claude Sonnet 5 launched on June 30, 2026 at $2 per million input tokens and $10 per million output — a third off the $3/$15 rate every Sonnet release had carried since Sonnet 4. Source: Anthropic, "Introducing Claude Sonnet 5" Anthropic was explicit that the discount was temporary: standard pricing of $3/$15 was scheduled to take effect September 1. We wrote about that countdown on July 25, when it was 37 days out and looked like the kind of pricing cliff worth planning around.

It didn't happen. Anthropic's own pricing documentation now carries this note: "The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur." Source: Anthropic, Claude Platform Docs, Pricing The cliff got cancelled seven weeks after it was announced, three weeks before it would have taken effect.

Pricing figures in this post are drawn from Anthropic's and OpenAI's published documentation as of August 2026, not from proprietary Nadir usage data.

We need to correct our own chart.

Our July post ran a three-bar chart: Sonnet 4.6 at $0.060 for a representative task, Sonnet 5 at $0.052 through August, and a red bar at $0.078 from September 1 — a 30% jump back over the old rate, driven by the discount expiring on top of Sonnet 5's newer tokenizer producing roughly 30% more tokens for the same text. Source: Anthropic, Claude Platform Docs, Pricing That red bar was accurate given what Anthropic had published at the time. It's also the one part of that post that turned out to be wrong, because the input it depended on changed.

Chart 1: The bar that never shipped. A representative 10,000-input/2,000-output-token task on Claude Sonnet 4.6 cost $0.060. The rate Anthropic announced for September 1, 2026 would have priced the same task at $0.078 on Sonnet 5 — that increase was cancelled August 10. Sonnet 5 has cost $0.052 for the same task since launch and still does today, now as the permanent rate.

That's the whole story in one image: the bar that should be there today, by Anthropic's own July roadmap, simply isn't. Teams that read the July pricing page and planned for $0.078 in September are budgeting, and possibly routing traffic, against a number that no longer exists.

The reversal didn't happen in a vacuum.

Six weeks separate the launch discount from the cancellation, and one of the loudest pricing moves of the year happened right in the middle of them. On July 30, OpenAI cut GPT-5.6 Luna's price 80% (from $1/$6 to $0.20/$1.20 per million input/output tokens) and Terra's 20%, just three weeks after that model family launched. Source: CNBC, "OpenAI cuts prices for two of its GPT-5.6 AI models as companies grow sensitive to costs," Jul 30, 2026; Source: OpenAI, "Advancing the price-performance frontier with GPT-5.6" OpenAI framed it as passing on efficiency gains; reporting at the time pointed at the same pressure showing up everywhere else in 2026: a more cost-sensitive enterprise buyer and competition from cheaper Chinese labs squeezing the exact price band Sonnet 5 competes in. Source: CNBC, Jul 30, 2026

Chart 2: Six weeks, four pricing decisions. Jun 30: Sonnet 5 launches at $2/$10, introductory through Aug 31. Jul 30: OpenAI cuts GPT-5.6 Luna 80% and Terra 20%, three weeks after that family's own launch. Aug 10: Anthropic makes $2/$10 the permanent Sonnet 5 rate. Sep 1: the $3/$15 hike that was announced at launch never happens.

Zoom out further and the pattern holds. The frontier token price index sits at roughly 12 as of late August 2026, 88% below its March 2023 baseline, even as frontier-tier models keep getting more capable. Source: BenchLM.ai, "LLM API Pricing Trends & Updates," Aug 2026 We've covered the mechanism behind that divergence before: flagship prices hold or climb while the mid-tier, the band Sonnet 5 and GPT-5.6 Terra actually sell into, keeps getting undercut. DeepSeek's pricing has been dragging that same band down all year. Raising Sonnet 5 by 50% into that environment, on the day the discount happened to expire, would have been the opposite move from where the rest of the market was heading. Cancelling it wasn't generosity. It was reading the room.

This isn't the first time the menu moved mid-year, either.

Sonnet 5's price staying put is one instance of a pattern this blog keeps documenting under different names. Anthropic closed Priority Tier to new capacity commitments in 2026 after selling it as a purchasable option earlier in the year. Three different providers renamed or restructured a model within the same 72-hour window in July, breaking anything that matched on a model string. Now a headline price change, announced explicitly and seven weeks in advance, got cancelled with three weeks of runway left. None of these required a single line of model behavior to change. Every one of them broke something for a team that had encoded the old assumption somewhere static: a price constant in a cost dashboard, a routing rule set to flip models on September 1, a budget forecast built on the $0.078 number.

What actually ships this week.

Grep your codebase and spreadsheets for hardcoded per-token prices. If a number like 0.000002 or $2/$10 appears anywhere outside a config value fetched at request time, it's a liability the day a provider changes its mind, not just the day a model gets deprecated.
Undo any pre-emptive routing changes made in anticipation of the September hike. If a team shifted default traffic off Sonnet 5 ahead of time, expecting a 50% jump, that traffic is now paying a different model's price for no reason — the thing it was trying to avoid never happened.
Treat a provider's pricing page as a diffable source, not a one-time read. A calendar reminder for "the day the discount ends" is exactly the kind of static assumption that just got proven wrong; a scheduled check against the actual page catches a reversal like this one before a forecast or a routing rule quietly goes stale.
Separate "the model changed" from "the price changed" in your monitoring. Model routing tends to get watched for capability drift; price is easy to assume is fixed between deprecations. 2026 has shown pricing moves on a shorter, less predictable cycle than model releases do.

Where this fits next to routing.

A cost dashboard or router that hardcodes $2/$10 for Sonnet 5 was correct on June 30, wrong-in-a-good-way through August, and would have been wrong-in-a-bad-way from September 1 if the cliff had landed as scheduled — three different truths for one constant, inside ten weeks. Nadir prices every routed request against live provider rates instead of a number someone typed into a config file once, so a reversal like Anthropic's on August 10 changes what a request costs on the savings dashboard automatically, the same day it takes effect, without anyone needing to notice the docs page changed. The same mechanism that keeps a routing decision from committing to a model based on a stale price is the one that keeps a cost forecast from committing to a rate that a provider is free to walk back with a documentation edit.

Conclusion.

Nothing about Claude Sonnet 5 changed on August 10 except the number written on Anthropic's pricing page, and even that number went back to what it already was. The interesting part isn't the model, it's the seven-week round trip: a price change announced with unusual specificity, three weeks of a competitor's own price cuts landing in between, and a reversal that undid the entire premise of a "cliff" post this blog published a month earlier. Teams that hardcoded September 1 into a spreadsheet, a routing rule, or a budget model are now holding a number Anthropic itself says will never be charged. The fix isn't predicting the next reversal correctly. It's not needing to, because the system pricing your requests reads the actual rate instead of remembering one.

Data in this post reflects Anthropic's and OpenAI's published pricing documentation as of August 2026, not measured production traces or proprietary Nadir customer data. Sources: [Anthropic, "Introducing Claude Sonnet 5"](https://www.anthropic.com/news/claude-sonnet-5). [Anthropic, Claude Platform Docs, Pricing](https://platform.claude.com/docs/en/about-claude/pricing). [CNBC, "OpenAI cuts prices for two of its GPT-5.6 AI models as companies grow sensitive to costs," Jul 30, 2026](https://www.cnbc.com/2026/07/30/open-ai-price-cut-gpt.html). [OpenAI, "Advancing the price-performance frontier with GPT-5.6"](https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/). [BenchLM.ai, "LLM API Pricing Trends & Updates," Aug 2026](https://benchlm.ai/llm-pricing-trends).

More on pricing & models

What Nadir is

Nadir is an LLM router. Nadir sizes every prompt and routes it to the cheapest model that still clears your quality bar. A small decision model reads each prompt against your model menu and picks the lowest-cost model it expects to handle it. When the decision model is uncertain or unavailable, the request goes to the strongest model on the menu. The routing step makes no call to your model provider.

Nadir runs two ways. The decision API returns a model, reasoning-effort, cache, context, and policy recommendation without calling a model provider, beside the gateway you already run. That is how a shadow-mode evaluation works, and its projected savings stay advisory. The OpenAI compatible managed proxy executes the route, and migration is a two-line change: point the base URL at api.getnadir.com and set model to auto. On that path an optional verifier can score a complete non-streaming answer and escalate to a stronger model when it misses the configured bar. Streaming bypasses post-generation verification. BYOK is supported on every tier.

For coding agents, Nadir Auto is a terminal launcher for Claude Code or Codex that asks Nadir for the main-session model before each turn, with your configured model as the fallback and cost ceiling. Delegation integrations for Claude Code, Codex, or Cursor instead recommend a model tier for subagent work, and the agent decides. Either way inference runs on your own provider account, so Nadir holds no provider keys on that path.

What the numbers are, and what they are not

Nadir publishes each evaluation with its scope. On checkable code, run-check-escalate solved 392 of 395 common HumanEval and MBPP problems (99.2%), graded by running the canonical tests; that applies only to tasks with runnable deterministic tests. Nadir-Tumbler posts an arena_score of 72.3 on RouterArena's public scorer, 5th of 23 routers, which measures the routing decision on RouterArena's own model pool. A reference-assisted RouterBench evaluation over 11,420 held-out triples produced a 60% lower projected cost than always-Opus with about 98% retained quality and a 1.7% catastrophic-route rate. That experiment gave the verifier the expensive-model reference answer, which production does not have, so it is a research ceiling and not the deployed path.

None of these is a production guarantee, a universal savings rate, or a forecast for any particular workload. Customer savings are reported from measured execution against a declared baseline, and customer quality only from outcome-labelled traffic. Projected savings and realized savings are separate artifacts and are never blended.

Design-partner program

Three rungs, picked by risk appetite. Rung 0 Shadow runs advisory decision calls alongside live traffic and returns a projected receipt, with nothing in the request path changed. Rung 1 Hosted is the two-line swap on a production slice and returns a realized receipt. Rung 2 On-prem is a supervised six-week proof of concept inside the partner's VPC, where no prompt, response, or usage reaches Nadir. The commitments are the same at every rung. Apply for a rung directly: Rung 0 Shadow, Rung 1 Hosted, or Rung 2 On-prem. Not sure which fits? Start at getnadir.com/contact/?reason=design-partner.

Licensing

NadirClaw is the self-hosted core, source-available under the PolyForm Noncommercial License. Source-available is the correct label; NadirClaw is not open source. Nadir Route's hosted plan has no base fee and charges a variable fee only on measured savings from requests Nadir executed.