# Half the price of frontier

URL: https://www.thedeepfeed.ai/posts/2026-07-24-half-the-price-of-frontier/
Category: Models
Published: 2026-07-24
Author: the-deep-feed
Tags: anthropic, claude-opus-5, pricing, benchmarks, coding-agents, frontier-models, cost-per-task
Kind: deep

> Anthropic shipped Claude Opus 5 as near-frontier intelligence at half the cost of Fable 5. The launch reads as good news for buyers. Read from the P&L, it is a margin admission — the clearest signal yet that the race has moved from raw capability to cost-per-task, and that nobody expects customers to pay the old premium for the top of the curve.

## TL;DR

- **Anthropic shipped Claude Opus 5 on July 24**, pitched in its own words as near-Fable-5 frontier intelligence at roughly **half the price**. Pricing held at **$5 / $25 per million input/output tokens** — unchanged from prior Opus.
- **Artificial Analysis** puts it narrowly **#1 on its Intelligence Index** while costing about **26% less per task** than the model it ranks just below the top. SOTA claims on **Frontier-Bench** and **GDPval-AA** (coding and knowledge work).
- The tell: Opus 5 **deliberately stays behind Mythos 5 on cybersecurity** — Anthropic says it is *not* SOTA on risky dual-use capability. A frontier lab is now choosing where *not* to lead. (That ceiling is its own story — more July 25.)
- The contrarian read: cheaper frontier looks like a gift to buyers, but for Anthropic it is a **margin admission**. When your headline model competes on price with the leader instead of beating it, you are conceding customers won't pay the old premium for the [top of the capability curve](/posts/2026-07-03-google-prices-the-generative-media-floor/).
- Stack it against the [open-weight floor](/posts/2026-07-04-the-gate-and-the-giveaway/): you can't charge for capability the giveaway is commoditizing. The battleground has moved to cost-per-task.

Anthropic released Claude Opus 5 on July 24, and the pitch was unusually candid about what it was selling. Not the best model in the world — close to it. The company's own framing, echoed across the launch coverage, was that Opus 5 delivers intelligence approaching that of Fable 5, the model most reviewers still treat as the raw-capability leader, at roughly half the cost per task. [Artificial Analysis](https://artificialanalysis.ai/articles/opus-5) put a number on it: Opus 5 edges into first place on its Intelligence Index while costing about 26% less per task than the model it had been trailing. [CNBC's](https://www.cnbc.com/2026/07/24/anthropic-claude-opus-5-ai-fable-5-cost.html) headline landed the same beat — matches the rival, undercuts the price.

Read as a product launch, this is straightforwardly good news, and most of the coverage read it that way. A near-frontier model got cheaper. Buyers win. But a release is also a disclosure, and the interesting document here is not the benchmark table. It is what a company that spent three years selling itself as the capability leader is telling you when its newest headline model competes on price instead of beating the field on power. That is not a victory lap. It is a concession about what customers are willing to pay for — and it is the clearest signal yet that the frontier race has stopped being about who has the best model and started being about who can serve a good-enough one for the least money.

# The $5/$25 price held, the value moved

Start with the artifact, because the pricing is the whole story. Opus 5 launched at **$5 per million input tokens and $25 per million output tokens** — identical to the prior Opus generation, as [MarkTechPost](https://www.marktechpost.com/2026/07/24/meet-the-new-claude-opus-5-frontier-class-agentic-coding-and-computer-use-at-unchanged-opus-pricing/) noted in its launch write-up. The number didn't move. What moved is what you get for it: Anthropic claims state-of-the-art results on **Frontier-Bench** and on **GDPval-AA**, the benchmark that scores models on economically valuable coding and knowledge work rather than on quiz-style reasoning. [Vellum's](https://www.vellum.ai/blog/claude-opus-5-benchmarks-explained) breakdown of the numbers puts the agentic-coding and computer-use gains as the real deltas over the last Opus, which tracks with where Anthropic has concentrated its commercial bets.

| | Claude Opus 5 | Reference: Fable 5 |
|---|---|---|
| Positioning (Anthropic) | Near-frontier | Raw-capability leader |
| Input price (per M tokens) | $5 | — |
| Output price (per M tokens) | $25 | — |
| Artificial Analysis Intelligence Index | Narrowly #1 | Just below |
| Cost per task (AA) | ~26% lower | Baseline |
| Cyber / dual-use | Deliberately behind Mythos 5 | — |

The load-bearing phrase in Anthropic's own copy is the comparison to Fable 5. When a lab benchmarks its flagship against a competitor's flagship and leads with *cost*, it has told you which axis it thinks the sale turns on. Not "we built the smartest model." Rather: "we built a model as smart as the one you were going to buy, for half the bill." That is a value pitch, and value pitches are what you make when the premium pitch has stopped closing deals.

# The frontier race changed axes

For most of the current cycle, the competitive question was capability, and it was scored in the obvious way: whose model tops the reasoning leaderboards, whose wins the hardest evals, whose can one-shot the demo that breaks the internet for a weekend. Price was a footnote — the assumption baked into every enterprise contract was that the best model was worth a premium, because the marginal capability at the top of the curve was where the value lived.

Opus 5 is Anthropic conceding that assumption no longer holds — at least not for the workloads that actually generate revenue. The tell is in what the release optimizes for.

![A labeled axis-rotation schematic on cream paper, black ink line-art with one red accent. Two perpendicular arrows share an origin: a faded, crossed-out vertical arrow labeled OLD AXIS — RAW CAPABILITY, and a bold horizontal arrow labeled NEW AXIS — COST PER TASK drawn in red, with model tokens placed along it: Fable 5 at the high-cost end, Opus 5 nearer the low-cost end annotated NEAR-FRONTIER AT HALF THE SPEND, and Kimi K3 (open) further along marked n/d. A rotation glyph at the origin reads THE AXIS TURNED.](/post-images/2026-07-24-half-the-price-of-frontier/axes-changed.jpg) Anthropic didn't ship a model to reclaim the raw-capability crown from Fable 5. It shipped a model to win on **cost-per-task**, the metric [Artificial Analysis](https://artificialanalysis.ai/articles/opus-5) now foregrounds, and it priced accordingly. This is the same shift we tracked when [Google priced the generative-media floor](/posts/2026-07-03-google-prices-the-generative-media-floor/) in early July: the competitive action moved from "what can the model do that no other model can" to "what does it cost to do the thing everyone's model can already do well enough." When the frontier commoditizes at the top, the sale moves to the denominator.

It is worth showing exactly how much of that denominator is actually disclosed, because the gap between what gets claimed and what gets published is the story in miniature.

| Model | Input ($/M tokens) | Output ($/M tokens) | Cost per task (Artificial Analysis) |
|---|---|---|---|
| Claude Opus 5 | $5 | $25 | ~26% below Fable 5 |
| Fable 5 (prior frontier) | n/d | n/d | Baseline (AA reference) |
| Kimi K3 (open weight) | n/d | n/d | n/d |

![A labeled cost-decomposition schematic on cream paper, black ink line-art with one red accent. Two cost bars on a shared baseline compare cost per task: a full-height Fable 5 (prior frontier) bar labeled BASELINE, and a shorter Claude Opus 5 bar annotated MINUS 26 PERCENT, split into stacked segments INPUT 5 PER M and OUTPUT 25 PER M with a bracket. A red delta wedge between the two bar tops reads THE 26 PERCENT — PRICE HELD BUT VALUE MOVED. The left axis reads COST PER TASK.](/post-images/2026-07-24-half-the-price-of-frontier/cost-per-task.jpg)

Only one row is fully sourced. Anthropic and [MarkTechPost](https://www.marktechpost.com/2026/07/24/meet-the-new-claude-opus-5-frontier-class-agentic-coding-and-computer-use-at-unchanged-opus-pricing/) give Opus 5 at $5 input and $25 output per million tokens, and [Artificial Analysis](https://artificialanalysis.ai/articles/opus-5) reports the roughly 26% lower cost per task against the model Opus 5 ranks just below. The intermediate token math, the per-task counts that turn a per-token price into a per-task cost, is not published, so the arithmetic underneath the 26% stays a black box the reader has to take on faith. Fable 5's own posted rates and Kimi K3's per-token pricing sit outside the launch corpus entirely, which is why the honest cells read n/d rather than a number invented to fill them. The [open-weight floor](/posts/2026-07-16-kimi-k3-open-frontier-ceiling/) undercuts all of this from below, but it does so on economics no vendor here has committed to a public rate card.

And there is a reason cost-per-task, not benchmark rank, is the honest metric now. A benchmark score is a claim about a ceiling. Cost-per-task is a claim about a budget: how much it costs to get a unit of useful work out the other end of a real agentic loop, tokens and retries and tool calls included. That is the number a buyer running a coding agent at scale actually feels, and it is the number Anthropic chose to lead with. We argued back in June, [measuring agent scaffolds](/posts/2026-06-22-how-much-is-the-harness-worth-measuring-agent-scaffolds/), that the harness around a model was quietly eating the model's share of the value. Opus 5 is the pricing consequence of that argument: if the scaffold does the differentiating, the model underneath it is a cost line, and cost lines compete on cost.

# Cheaper frontier is a margin admission

Here is the reading the launch coverage mostly skipped. Cheaper frontier is unambiguously good for buyers. For Anthropic, it is an admission dressed as a feature.

Think about what "near-Fable-5 at half the price" does to Anthropic's own book. The premium tier of the AI market — the customers who will pay top dollar for the last 10% of capability, the frontier of the frontier — is the highest-margin revenue a lab can book. It is the segment that justifies the enormous training runs, the one where you get to charge for being *the best* rather than for being *good enough*. When Anthropic ships a model explicitly positioned to match the leader at half the cost, it is telling its own sales force that this premium segment is thinning — that a growing share of customers will not pay the differential for the top of the curve, and will instead take 95% of the capability at 50% of the price and never look back.

That is cannibalization by design. A model priced to win on cost-per-task competes hardest against your *own* highest-margin business, because the customers most likely to defect to "good enough at half price" are exactly the ones who were paying you the premium. Anthropic is choosing to lead that migration rather than let a competitor — or an open-weight release — lead it for them. Rational, probably necessary, and still a concession: you only race to commoditize your own frontier when you have concluded that someone is going to commoditize it regardless.

Who is that someone? The open-weight floor. When Meituan open-sourced a 1.6-trillion-parameter agentic-coding model under an MIT license — [the gate and the giveaway](/posts/2026-07-04-the-gate-and-the-giveaway/) — it set a price of zero on near-frontier capability for anyone willing to self-host. You cannot charge a premium for capability that a giveaway is driving toward free. The Chinese open-weight releases and Anthropic's price-halving are the same pressure read from opposite ends: one commoditizes capability by giving it away, the other pre-empts the commoditization by pricing its own frontier as if it had already happened. The premium on intelligence is collapsing from both directions at once.

# The one place Anthropic chose not to lead

There is a striking exception in the Opus 5 release, and it is not an accident. Anthropic states plainly that Opus 5 is **not** state-of-the-art on cybersecurity — that it deliberately trails Mythos 5 on the dual-use capabilities a security-tuned model is built to exercise. In a launch otherwise built around leading, the lab went out of its way to say where it chose *not* to.

That is a remarkable thing to volunteer, and it doesn't fit the cost-per-task story cleanly, which is why it deserves its own treatment. A lab racing purely on economics would push capability everywhere the market rewards it and let the chips fall. Anthropic instead drew a ceiling — a capability it could presumably ship but decided not to lead on, for reasons that are about governance and dual-use risk rather than margin. That decision turns "how good is the model" into "how good is the model *allowed to be, by its maker*," and it converts what we described in July as a [voluntary policy gate](/posts/2026-07-09-the-voluntary-gate-that-works-like-a-license/) into something closer to a product-line decision. Google is racing *toward* cyber capability in the same fortnight; Anthropic is deliberately racing *away* from it. That contrast is the subject of tomorrow's piece. For today it is enough to note that the same release that concedes the capability premium on coding also refuses the capability premium on cyber — one for economics, one for principle.

# How builders reacted

The discourse around the launch was, honestly, quiet. Opus 5 arrived without the leaderboard-war frenzy that greeted earlier frontier drops; the harvested window skews heavily toward builders posting demos — one-shot landing pages, procedurally generated games, side-by-side dev streams — rather than toward anyone litigating the pricing thesis. Engagement was modest across the board; the top on-topic posts sat in the low tens of likes, not the thousands. That relative silence is itself a signal: a near-frontier model at half price now lands as a routine product update, not an event. The premium is gone from the *reaction*, too.

The take that most cleanly matched Anthropic's own framing came from a developer-tools account:

> Claude Opus 5 is getting dangerously close to Fable 5. The interesting part? It costs half as much.

> — [@appwrite](https://x.com/appwrite/status/2082056449879113899), Jul 28 · 3 likes, 2 bookmarks, ~456 impressions

The skeptics landed on the other half of the thesis — that "close to frontier" is doing a lot of work, and that the benchmark gap you can't see in a leaderboard shows up in the loop:

> opus 5 is not better than fable. stop blindly believing the benchmarks

> — [@LukeberryPi](https://x.com/LukeberryPi/status/2082062150600634653), Jul 28 · 3 likes

> I keep seeing people posting that Opus 5 made them a video game with one prompt while mine is struggling with figuring out if it made a PR to the repo or not. It's a poor man's Fable.

> — [@filrakowski](https://x.com/filrakowski/status/2082054299455308154), Jul 28 · 2 bookmarks

"A poor man's Fable" is, unintentionally, the whole product strategy in four words — and whether that reads as an insult or a value proposition depends entirely on whether you were ever going to pay the rich man's price. A third strand doubted the release-day numbers survive contact with production at all, one user calling Opus 5 *"a worse model than opus 4.7 at this point"* despite the launch benchmarks ([@i_am_vickyd](https://x.com/i_am_vickyd/status/2082059358020801015), Jul 28). None of it out-traveled the demo videos. The market's verdict on cheaper frontier was a shrug — which is exactly what a collapsing premium sounds like.

# The premium was the moat

> **The Deep Feed's position:** the launch coverage read "near-Fable-5 at half the price" as a gift to buyers. It is that — and it is also Anthropic telling its own sales force the premium segment is thinning. When your number-two pricing deliberately cannibalizes your number-one capability, you are pricing as though the last 10% of intelligence is no longer something customers will pay a differential for. We read Opus 5 not as a discount but as a confession, and the confession is the same one the whole fortnight keeps making: the capability premium is going, and cost-per-task is the metric that replaced it.

For three years the implicit business model of every frontier lab was that intelligence carried a premium — that being the best was worth being the most expensive, and that the customers who needed the top of the curve would pay for it. Opus 5 is Anthropic pricing as though that premium is already gone: matching the leader instead of beating it, holding the price flat while doubling the value, and leading the migration to good-enough-at-half-cost before someone else — an open-weight download that asks no one's permission — leads it for them.

The optimistic headline is that frontier AI got cheaper, and it did. The sharper reading is that a company which spent years selling capability just told the market it no longer expects to be paid for the last 10% of it. When your #2 pricing cannibalizes your #1 capability on purpose, the moat you're admitting you no longer have is the one every lab was built on. What's left to compete on is the margin — and that is where this fortnight has been heading all along.

## Sources

- [Anthropic — Introducing Claude Opus 5 (Jul 24, 2026)](https://www.anthropic.com/news/claude-opus-5)
- [Artificial Analysis — Claude Opus 5: Fable 5 level intelligence at lower cost per task (Jul 24, 2026)](https://artificialanalysis.ai/articles/opus-5)
- [VentureBeat — Anthropic launches Claude Opus 5, a cheaper AI model for coding agents and enterprise workflows (Jul 24, 2026)](https://venturebeat.com/orchestration/anthropic-launches-claude-opus-5-a-cheaper-ai-model-for-coding-agents-and-enterprise-workflows)
- [CNBC — Anthropic's Claude Opus 5 matches rival Fable 5 at lower cost (Jul 24, 2026)](https://www.cnbc.com/2026/07/24/anthropic-claude-opus-5-ai-fable-5-cost.html)
- [Vellum — Claude Opus 5 benchmarks explained (Jul 24, 2026)](https://www.vellum.ai/blog/claude-opus-5-benchmarks-explained)
- [MarkTechPost — Meet the new Claude Opus 5: frontier-class agentic coding and computer use at unchanged Opus pricing (Jul 24, 2026)](https://www.marktechpost.com/2026/07/24/meet-the-new-claude-opus-5-frontier-class-agentic-coding-and-computer-use-at-unchanged-opus-pricing/)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-07-24-half-the-price-of-frontier/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt