Anthropic shipped Claude Opus 5 into general availability on July 24, 2026. It costs $5 per million input tokens and $25 per million output tokens. That is exactly half of Fable 5's $10 input and $50 output, and identical to what Opus 4.8 cost since May 28, 2026.
Here is the part that should stop you. On CursorBench 3.2 at max effort, Opus 5 lands within 0.5% of Fable 5's peak score, at half the cost per task. On Frontier-Bench v0.1 it scores 43.3%, more than doubling Opus 4.8's 21.1%.
Anthropic did not lead with a parameter count. The company did not publish one. The launch post led with price, effort settings, and the phrase "designed to be used every day."
The timing was pointed. Opus 5 arrived within days of OpenAI publishing Astra results on 10 unsolved math and theory problems. The competitive response cycle in frontier AI is now under a week.
Two labs, two different pitches, same news cycle. One said: we advanced human knowledge. The other said: we made the thing you already run in production twice as good for the same money. I think the second pitch is the one that moves enterprise budgets this quarter.
The Workhorse Rule
The model that wins the enterprise is not the smartest one. It is the one you can leave running.
Four numbers that explain why the flagship price sheet did not move.
That is the Workhorse Rule, and it explains nearly every choice in the Opus 5 release. Anthropic's own documentation calls it a model "for complex agentic coding and enterprise work." Not for solving open problems in number theory. For work that runs long, runs often, and has to not fall over.
Sort your workloads into three buckets and the tier structure snaps into focus.
Reflex work. Classification, extraction, routing, tagging. High volume, low stakes, per-token cost dominates. Haiku territory.
Range work. Drafting, summarizing, everyday coding. Sonnet 5 shipped June 30, 2026 at $2/$10 through August 31, then $3/$15. Good enough, cheap enough, most of the day.
Run work. Multi-file refactors, week-long agent sessions, migrations across a 40-file codebase. Failure here is expensive, and the cost of a wrong answer dwarfs the cost of tokens. Opus tier.
Frontier benchmark scores tell you about ceiling. The Workhorse Rule is about floor. Enterprise buyers are not asking which model is smartest on its best day. They are asking which model does not silently break the Tuesday night data migration.
Under the Hood: Why Thinking-On-By-Default Changes Your Build
Let me show you what actually shifted, because the benchmark table is not the interesting part.
Opus 5 has a 1 million token context window. That is both the default and the maximum, with no smaller variant. Max output is 128,000 tokens. Thinking is on by default, which breaks the way you built against Opus 4.8.
Effort is a dial now: low, medium, high, xhigh, max. Same model, five prices. On Frontier-Bench and CursorBench, Anthropic published performance-per-cost curves at each setting rather than a single hero number. Translation: they handed you a throttle and expect you to use it.
Napkin math on what that means for your wallet. Feed an agent a 200,000-token repo snapshot and let it write 20,000 tokens of patch. Input costs $1.00. Output costs $0.50. Call it $1.50 a run, ten runs a day, roughly $450 a month for one agent that works while you sleep.
Now the two features nobody put in a headline.
Mid-conversation tool changes, in beta behind the mid-conversation-tool-changes-2026-07-01 header. You can add or remove tools between turns while preserving the prompt cache. Before, changing an agent's toolset meant resending a fixed list and eating the cache miss. Now your agent can pick up a database tool at step four and drop it at step nine without paying twice.
Default fallbacks mode. The fallbacks parameter accepts "default", which applies Anthropic's recommended fallback models by refusal category. You stop hand-maintaining a list of backup models for every edge case. Small feature. Deletes an entire class of glue code.
An AI agent is a 500 IQ intern with amnesia. Every capability that survives across turns, cached prompts, persistent tools, cloud-hosted sessions, is a patch on the amnesia. That is what Claude Cowork moving to Anthropic's cloud on August 3, 2026 is really about. Cowork used to run on one laptop. Now a session spans devices and time, on by default for teams, manageable from organization settings.
A model that only wins benchmarks is a Ferrari: gorgeous, no engine mounted in anything you own. Now look at where Opus 5 is wired. Bedrock across US East, Melbourne, Ireland and Stockholm. Google Cloud's agent platform. Microsoft Foundry. A Microsoft 365 connector that can now write and not just read. That is a tractor. Tractors do the farming.
Simple always defeats complex. The 20% here: pick your effort level per task type, cache your prompts, and stop rebuilding fallback logic by hand.
Three signals inside the same shift
The flagship got better and the invoice did not move.
Opus 5 holds the same $5 input and $25 output pricing Opus 4.8 carried since May 28, 2026, while more than doubling Frontier-Bench performance to 43.3%. That is half of Fable 5's $10 and $50. The tokens are the loss leader; the workflow is the product.
Fable 5 went generally available and then vanished.
Anthropic made Fable 5 GA on June 9, 2026. On June 12 a US government directive forced suspension of all access for any foreign national, with a compliance window measured in minutes. The Cloud Security Alliance now treats single-vendor frontier dependence as demonstrated risk, not hypothetical.
Unsolved conjectures and closed tickets are not the same skill.
OpenAI published Astra results on 10 unsolved math and theory problems within days of the Opus 5 launch. No one has shown that capability transfers to enterprise workflow depth. Most companies only get paid on one of these two scoreboards.
2031
Zoom out five years and the price sheet matters more than the leaderboard.
Costco has sold the same hot dog and soda for $1.50 since 1985. The hot dog is not the business. The membership is the business. Anthropic holding Opus at $5/$25 for two consecutive flagship releases, while shipping a model that more than doubles Frontier-Bench performance, is the same move. The tokens are the hot dog. The workflow is the membership.
Benchmarks buy headlines. Workflows buy renewals.
Now the asymmetric risk, and it is real. On June 9, 2026, Anthropic made Fable 5 generally available. On June 12, three days later, a US government directive forced Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national, including its own foreign employees, with a compliance window measured in minutes. That order restricted access. It did not protect customers who had already built on it.
Anyone who had wired Fable 5 into a production workflow learned the lesson in 72 hours. Deep integration with a single frontier provider is a switching-cost moat that points both directions. The Cloud Security Alliance has been blunt that concentrated dependence on one frontier vendor is now a demonstrated risk, not a hypothetical one.
So the strategic read is split. Opus 5 became the highest generally available Claude model partly because the tier above it got locked down. Anthropic's own system card says Opus 5 is not more capable overall than Fable 5, and that it remains behind Mythos 5 on cybersecurity tasks. A deliberately non-frontier flagship is a strange thing to call an inflection point, and the security analysts who think Mythos was the real turn have a case.
And OpenAI's Astra results on 10 open math and theory problems are a different kind of claim entirely. It is unclear whether enterprise buyers can operationalize that at all in 2026. Solving an unsolved conjecture and closing 400 Jira tickets are not the same skill, and no one has shown the transfer yet.
My read on this: the labs are now competing on two separate scoreboards, and most companies only get paid on one of them. Build for the one that ships.
What to Build This Weekend
Pick one workflow. Not five. Ship it by Sunday.
Step 1: audit your default model. Open whatever agent or script you already run. If it hardcodes an older Claude ID, note where. Opus 5 has thinking on by default, so test before you swap. Things will break. That is normal, and finding out on a Saturday beats finding out on a Monday.
Step 2: build one long-horizon job. OpenAI Codex just went generally available with adaptive thinking and long-running refactor jobs. Give it something genuinely multi-hour: dependency upgrade across a repo, test coverage on a neglected module. Then compare that same task against Opus 5 at high effort and again at max. Log the cost per task. You now have your own performance-per-dollar curve instead of a vendor's.
Step 3: give your agent eyes. Chrome DevTools MCP 1.5.0 added memory analysis to its Model Context Protocol server. MCP is a standard plug that lets a model talk to an outside tool without custom code. With 1.5.0, an agent can take a heap snapshot and reason about a memory leak it could previously only guess at. Wire it up and ask it why your page slows down after five minutes.
Step 4: chain two models, not one. Brainy Canvas lets you build workflows spanning multiple models on a single canvas. Route cheap classification to a small model and hand only the hard step to Opus 5. That is the Workhorse Rule as an architecture diagram.
Step 5: automate one annoying thing. Flick prioritizes incoming email and drafts candidate replies. Install it, let it run for a week, and count how many drafts you send unedited. If the number is above half, you have proof. If it is below, you have a prompt to fix.
No CS degree required for any of this. Get your reps in, break things in a sandbox, and keep the receipts on cost per task. The teams that will win 2027 are the ones measuring that number today.
Ship one long-horizon job and measure cost per task.
- Audit your default model ID. Open the agent or script you already run in production and find every hardcoded Claude version. Opus 5 has thinking on by default, so test the swap in a sandbox before Monday, because things will break.
- Build your own performance-per-dollar curve. Give one genuinely multi-hour task, like a dependency upgrade across a repo, to Opus 5 at high effort and again at max, then log cost per task. A 200,000-token repo snapshot plus 20,000 tokens of patch runs roughly $1.50 per run, about $450 a month at ten runs a day.
- Give the agent eyes and a cheap first stage. Wire up Chrome DevTools MCP 1.5.0 so your agent can take a heap snapshot and reason about a real memory leak, then route classification to a small model and hand only the hard step to Opus 5.
The winning model is not the smartest one. It is the one you can leave running.
Anthropic did not publish a parameter count and did not claim Opus 5 beats Fable 5 overall; the system card concedes it trails Mythos 5 on cybersecurity tasks. What it claims instead is that a model you already run in production got more than twice as good on Frontier-Bench at the same $5 and $25 price, with a 1 million token window, five effort settings, and deployment across Bedrock, Google Cloud and Microsoft Foundry. That is a workflow bet, not a benchmark bet, and the Fable 5 suspension three days after GA is the reason hedging matters. Sort your workloads into reflex, range and run, then price each one. The teams that win 2027 are measuring cost per task today.