A New Flagship-Class Model Shipped With an Unchanged Price Tag

Anthropic released Claude Opus 5 on July 24, 2026, and buried in the announcement is a detail most coverage skipped past. The price is $5 per million input tokens and $25 per million output tokens — with a parenthetical noting it's "the same as Opus 4.8." Not a cent of difference from the previous generation, which shipped on May 28, less than two months earlier. Yet a big share of the headlines that same day carried the word "cheaper" or "half the cost."

Here's the deal: both statements are true, and understanding why is the entry fee for reading this launch correctly. "Half" is measured against Claude Fable 5, Anthropic's top-end flagship, which runs $10 input and $50 output. Opus 5 is positioned to deliver something close to Fable 5 quality at half of Fable 5's rate. That's where the number comes from. If somebody tells you "Opus got cheaper this week," that sentence is wrong. The Opus tier has been pinned at $5/$25 since the 4.5 generation, and it's still there. What changed is how much capability fits inside that price, not the price.

The distinction matters because it completely changes the math a company runs when planning an AI budget. If you read this as "the price dropped 50%," you expect your existing Opus invoice to get cut in half. It won't. The per-token rate is identical. What actually happens is different and slower: the model finishes tasks with fewer wasted round trips, so tokens burned per task falls, and any workload you were routing to Fable 5 that Opus 5 can now handle drops to half its previous rate. That's the whole mechanism. It's also why Anthropic never states a benchmark score in this launch without attaching cost per task to it.

And this didn't happen in a vacuum. Anthropic is a company that reportedly has an IPO in motion, and its rivals are shipping new frontier models on a multi-week cadence. In the middle of that, the card Anthropic chose to play was: leave the price alone, push the capability. That's the real story here.

Anthropic, Fable, and What the 'Opus' Slot Is For

Anthropic is an AI lab founded in 2021 by a group of former OpenAI researchers. Google and Amazon are among its major investors, and it ships a model family called Claude. From day one the company has built its public identity around safety, and that flavor still shows up all over the model cards and launch posts. In the Opus 5 announcement, right after the performance claims, Anthropic goes out of its way to spell out where the model's capabilities stop on cybersecurity and biology tasks. That's the house style.

The Claude lineup has gotten genuinely layered. At the top sits Claude Fable 5, generally available since June 9, which Anthropic's own docs describe as its most capable widely released model. It runs $10 input, $50 output. Next to it is Claude Mythos 5, a defensive cybersecurity model available only to invited customers, so most developers will never touch it. Below that is the Opus line. Below that, Sonnet 5 and Haiku 4.5 handle the speed-and-price axis.

So "Opus" isn't the peak. Anthropic's developer documentation describes Opus 5 as the model for complex agentic coding and enterprise work, and tells you to start here if you're unsure which model to pick. The very next sentence says to move up to Fable 5 if you need maximum capability. The division of labor is clear: Opus is the daily driver, Fable is the thing you pull out when a problem is genuinely hard. This launch is about dragging the daily driver up close to the flagship.

Understanding that layout also explains why the price didn't move. Push Opus 5 below $5/$25 and it collides with Sonnet 5 directly underneath it ($3 input / $15 output, with a $2/$10 introductory rate running through August 31). Above it, the 2x gap to Fable 5 is what makes the ladder legible in the first place. So $5/$25 isn't inertia. It's the rung Opus is supposed to occupy on a four-tier price ladder, and moving it would knock the other rungs loose.

One more piece of background. There have been persistent reports that Anthropic has entered IPO preparation — a confidential filing with the SEC in June, with a Nasdaq debut targeted for the fall. The company's position is that it hasn't decided on timing or on whether to go public at all. So treat that as reporting and speculation, not settled fact. But whether or not it pans out, it's worth holding onto the context: at a moment when a company would need to show revenue growth and margin structure at the same time, the product it shipped holds price flat and raises capability per dollar.

The Benchmarks Are Stated in Cost per Task, Not Just Score

Every headline metric in this launch pairs performance with cost. Pulling the phrasing straight from the official announcement: on Frontier Bench v0.1, Opus 5 beats every other model and more than doubles Opus 4.8's result, at lower cost per task. On CursorBench 3.2 at maximum effort, it lands within 0.5% of Fable 5's best score at half the cost per task. On ARC-AGI 3, it scores three times the next-best model. On Zapier AutomationBench, at equal cost per task, its pass rate is roughly 1.5x the runner-up. On OSWorld 2.0, it exceeds Fable 5's best result at roughly a third of the cost.

The specific numbers came from reporting. On Frontier Bench v0.1, Opus 5 reportedly scored 43.3%, Fable 5 33.7%, and OpenAI's GPT-5.6 Sol 34.4%. That benchmark comes from the team behind Terminal Bench and Harbor, and it measures the share of multi-step engineering tasks an agent carries all the way to completion. One number to watch skeptically: Opus 4.8's score. Different outlets put it anywhere from 18.7% to 21.1%, and it's too early to say which is right. Both figures are at least roughly consistent with Anthropic's own "more than double" phrasing, so nothing is obviously broken — but the spread is real.

The change most likely to show up in daily work isn't a benchmark score at all. It's the effort toggle. Per Fortune's reporting, users can choose per request how hard the model works — low, medium, or high. Anthropic's docs note that Opus 5 defaults to high in the API and in Claude Code. For a company watching token spend, this is finally a knob for the complaint "why is it thinking at full tilt on a request that didn't need it?" Separately, there's a fast mode running roughly 2.5x the base speed, priced at double the base rate: $10 input, $50 output.

The safety writeup is worth reading too. Anthropic says automated behavioral audits gave Opus 5 the lowest misaligned-behavior score of any of its recent models, while explicitly stating the model still falls short of Mythos 5 on cybersecurity vulnerability exploitation and biology research tasks. The claim that it doesn't increase dual-use risk — capability that could be turned to offensive or military purposes — is in there for the same reason. The argument is that the capability curve went up while the risk curve didn't, and that reads as a message aimed at regulators and large enterprise buyers as much as at developers. TechCrunch reported that, unlike Fable and Mythos, this model doesn't carry a 30-day data retention policy, and that it now automatically reroutes requests caught by safety filters to a weaker model instead of refusing outright.

Item Claude Opus 5 Claude Fable 5 Claude Opus 4.8
General availability 2026-07-24 2026-06-09 2026-05-28
Model ID claude-opus-5 claude-fable-5 claude-opus-4-8
Input price (per 1M tokens) $5 $10 $5
Output price (per 1M tokens) $25 $50 $25
Fast mode $10 / $50, ~2.5x base speed — —
Context window 1M tokens 1M tokens 1M tokens
Max output 128K tokens 128K tokens 128K tokens
Reliable knowledge cutoff May 2026 Jan 2026 Jan 2026
Frontier Bench v0.1 (reported) 43.3% 33.7% 18.7–21.1%, sources disagree
Positioning (per official docs) Complex agentic coding, enterprise work Next-gen intelligence for long-running agents Filed under legacy models

One more thing the table surfaces: Opus 4.8 moved into the "legacy models" accordion in Anthropic's official docs two months after launch. That's a reasonable bit of housekeeping once a same-priced successor exists. But two months is not a comfortable interval for anyone responsible for enterprise adoption. Plenty of organizations take longer than that just to validate a model and get it into production.

What Each Side Gets Out of This

Anthropic's side of the ledger is straightforward. Price held, so per-token margin structure holds. It gets to say "twice the capability for the same money." And it can absorb high-difficulty workloads that would otherwise have gone to Fable 5 into a model that costs half as much to run for the customer. Which creates an interesting paradox: every workload Opus 5 pulls down from Fable 5 halves the revenue on that workload. And because Opus 5 finishes tasks in fewer round trips, tokens per task drop too. On the surface, Anthropic looks like it's cannibalizing itself. That math only works if volume grows — existing users running more because the unit rate is lower, plus workloads that were previously priced out entering the market at all. That's the bet.

For developers and engineering teams, the win is predictability. The most annoying part of swapping models is usually that the entire cost model changes underneath you. Here the price sheet is identical, so nobody has to rebuild a budget spreadsheet, and the migration decision collapses down to one question: did quality actually improve? Add the effort toggle and you can now pick a cost-performance point per workload. High-frequency, low-difficulty calls like autocomplete go to low. Overnight batch refactors go to high. That's real operational control that didn't exist before.

For enterprise buyers the benefit sits on a different layer. CNBC put the cost-sensitive enterprise angle right in its headline, and that's the correct read. The recurring pattern across the last year of AI adoption projects has been: the pilot works, then the company-wide rollout stalls on token spend. An option with Fable-class output at Opus-class pricing unsticks some of those. Per reporting, Anthropic compared Opus 5 and Fable 5 across 13 benchmarks and Opus 5 came out ahead on 8. Flip that around: Fable 5 still wins on 5. The top model didn't become unnecessary. The range where you need it just got narrower.

Now the risks, stated honestly. First, benchmark conditions may not match your conditions. That 0.5% CursorBench gap came from the maximum effort setting, and maximum effort means burning a lot of tokens. Whether the same gap survives at default settings is something you have to measure on your own workload. Second, a two-month model cadence generates recurring validation cost. A team that just finished getting Opus 4.8 approved and deployed is now facing the same process again eight weeks later. Third, the effort toggle is as much a management problem as a feature. Let several teams inside one organization each set their own effort levels, and tracing the cause of a cost spike three months later gets genuinely hard.

Precedents — What Worked and What Didn't

"Raise capability, hold or cut price" is a strategy this industry has already run several times, with mixed results. The success case people cite most is OpenAI's GPT-4o. Released in May 2024, it halved API pricing versus the incumbent GPT-4 Turbo and roughly doubled throughput. The effect was that a large population of applications stuck on GPT-3.5 purely for cost reasons moved up a tier, and OpenAI grew total usage faster than it cut unit price. That's the textbook case where a price cut expanded the market instead of shrinking revenue.

Anthropic has its own precedent for this exact move. When Claude 3.5 Sonnet launched in June 2024, Anthropic kept Sonnet-tier pricing ($3 input / $15 output) unchanged while claiming the model outperformed Claude 3 Opus — then its top model at $15 input / $75 output — across multiple evaluations. A mid-tier model at the same price beat a flagship costing five times more. Claude's standing in the developer tooling ecosystem jumped hard after that, and it's the moment "Anthropic is good at coding" became conventional wisdom. The structure of the Opus 5 launch is nearly a carbon copy: price fixed, capability eating into the tier above, and cost per task as the framing device.

The counterexample is instructive. OpenAI's GPT-4.5, released as a research preview in February 2025, arrived at $75 per million input tokens and $150 output — extreme even by frontier standards. Reviewers found the model genuinely impressive, but it never escaped the argument about whether it was worth the money. In under two months OpenAI announced it would pull the model from the API, and it was actually removed in mid-July that year. The lesson: a model that doesn't pencil out on cost per task doesn't reach production, no matter how good it is. Anthropic attaching a cost figure to every performance claim in this launch looks like a direct application of that lesson.

There's a second failure mode worth naming: benchmark trust. In April 2025, Meta took heavy criticism over Llama 4 when it emerged that the version it submitted to a competitive leaderboard was an experimental build different from the one shipped publicly. Once announced numbers and the model users actually receive diverge, every subsequent number from that vendor gets discounted. Opus 5's figures have to clear the same bar. Whether the 0.5% CursorBench gap, the 3x on ARC-AGI 3, and the one-third cost on OSWorld 2.0 hold up under independent reproduction is something the next few weeks will settle.

How Rivals Counter

OpenAI enters this round with GPT-5.6 Sol at 34.4% on Frontier Bench — ahead of Fable 5's 33.7%, but nearly nine points behind Opus 5's 43.3%. OpenAI's standard playbook has two moves. One is shipping a stronger model fast. The other is repricing an existing model or expanding reasoning-budget options to redraw its own cost-per-task curve. GPT-4o proved OpenAI will actually pull the price lever, so a counterattack aimed squarely at the "half the price" narrative isn't a stretch. The complication is that OpenAI's own public listing is reportedly further out than Anthropic's, which makes the question of how aggressively it can afford to compress margin a separate matter entirely.

Google fights on a different axis. Its advantage has never really been raw benchmark position — it's distribution. Cloud, Search, Workspace, Android: the sockets are already installed. Google can be a generation behind on a leaderboard and still not lose ground at the point of user contact. It's also complicated: Google is a major Anthropic investor, and Claude models are served on Google Cloud. Being simultaneously a competitor and a distribution channel blunts any move Google might make. The Google-shaped response isn't a public price war — it's quietly lowering unit cost through its own silicon and letting that compound.

The sharpest contrast is xAI. It shipped Grok 4.5, tuned for coding and agentic work, on July 8, and on July 25 announced via its official account that the model was now wired into its Augment feature. That's the day after Opus 5 launched. xAI has the shortest ship-to-deploy cycle among the frontier labs, and it runs a structure where a new model goes straight into its own products so usage data comes back immediately. Where Anthropic played a static card — hold the price, raise capability density — xAI is playing a dynamic one: ship fast, wire it into the product, spin the learning loop.

These two strategies compete on different surfaces, so there's no fast verdict. Anthropic's approach is built for enterprise procurement. A stable price sheet and a thick safety document make it much easier to clear sourcing, legal, and security review. xAI's approach is better at rapidly capturing daily usage among individual developers and small teams. The catch is that those two markets eventually converge. The path where a tool a developer adopted personally ends up inside the company has repeated itself over and over these last few years.

There's also pricing pressure from Chinese open-weight models that can't be waved off. The set of options that get close to frontier capability while charging single-digit cents keeps growing, and that pulls the floor down under every closed model's price. How long Anthropic can hold the Opus line at $5/$25 depends on how long it can offset that pressure with a capability gap.

So What Actually Changes

If you're a developer, the first thing to do isn't recalculating costs — it's classifying workloads. Measure how much of what you currently run on Fable 5 holds up when moved down to Opus 5. Everything in that range immediately costs half of what it did. Going the other direction, moving from Opus 4.8 to Opus 5 involves zero price change, so the decision is a comparatively simple one: check for regressions and ship. And don't leave the effort toggle at defaults. Both the API and Claude Code default to high, so if you configure nothing, trivial requests get the full treatment and the full token bill.

If you're on the enterprise side, the first question is whether your process can survive a two-month model cadence. Opus 4.8 was filed under legacy in the official docs eight weeks after launch. If your model approval cycle runs quarterly, you'll repeatedly finish approving a model right as it becomes previous-generation. The practical fix is to stop approving individual models and start approving at the vendor-and-price-tier level, with swap-in rights inside that tier.

If you're looking at this as an investor, the number to watch isn't price — it's volume. Anthropic defended its unit rate, but every Fable 5 workload that migrates down to Opus 5 reduces revenue on that workload. For the structure to be net positive, total usage has to grow by more than that decline. If the IPO reporting is accurate, that's exactly what the disclosed financials would need to show: not token pricing, but processed volume and usage per customer over time. Since the company hasn't confirmed timing or even the decision to list, it's too early to plan around a date.

If you're a regular user, the change arrives quietly. Opus 5 is now the default model on Claude Max and the top model on Claude Pro, and it's available in Claude Code and Claude Cowork. Subscription pricing didn't change — you're just getting a better model at the same rate. Per Anthropic's docs, the context window is 1 million tokens and max output is 128K tokens. The reliable knowledge cutoff is May 2026, four months newer than Fable 5's January 2026. For work that touches recent events, that gap will be more noticeable than the benchmark numbers.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? If you're on Claude Pro or Max, it already changed something: your default model became Opus 5 at no extra charge. If you're an API customer, the price sheet didn't move, so there's no budget rework — just check for quality regressions before you switch.

— Is the "half price" claim just marketing spin? It's not false, it's a different baseline. Against Fable 5 at $10/$50, half is exactly right. But if you're coming from Opus 4.8, your price change is precisely zero, and if you don't check which model a "half price" claim is measured against, your budget planning goes sideways.

— Should I trust the benchmark numbers as published? Hold off for now. The 0.5% CursorBench gap comes from the maximum effort setting, and Opus 4.8's Frontier Bench score is reported anywhere from 18.7% to 21.1% depending on the outlet. Too early to call until a few weeks of independent reproduction pile up.

References

Numbers are as of announcement and may change.