Claude Opus 5 Launches at Half the Price of Fable 5, and Anthropic's Efficiency Claims Have a Track Record Problem
Anthropic shipped Claude Opus 5 today: near-Fable 5 scores, a five-level effort dial, half the price. History says verify the token math before you trust it.
Two prompts. That's how much work I get out of Claude's free tier on a bad day before it locks me out for four hours. I've caught myself pricing out ChatGPT and Gemini over it more than once. The meter runs out before the job does, on a loop, week after week.
Today Anthropic shipped Claude Opus 5, its fourth model release in under two months, and the pitch is a five-word promise: use fewer tokens to get there. Anthropic frames Opus 5 as a model that "comes close" to the frontier intelligence of Claude Fable 5 at half the price, built for daily use instead of benchmark ceilings.
Token Efficiency, Defined
A token is the unit a language model reads and writes, not a word. English text runs somewhere around a word and a half per token, and that ratio shifts depending on the tokenizer, the specific vocabulary Anthropic trained the model to chop text into. Two consequences follow, and both matter for the rest of this article.
First, every AI lab bills by the token, so a model that reaches the same answer using fewer of them costs less no matter what the price-per-token sticker says. Second, the tokenizer itself can change between model versions, and the exact same English sentence can then map to a different token count, even though the words on the page never moved.
Opus 5 adds a second lever on top of the tokenizer: an API field called effort. Anthropic exposes five settings, low, medium, high, xhigh, and max, that let a single model trade thoroughness for token spend on a per-request basis. Instead of switching models to save money, a developer sets one parameter. Anthropic's prompting guide for Opus 5 recommends low and medium for most requests, and xhigh as the starting point for coding and agentic work.
The same dial sits in the consumer app too, next to the model picker in Claude.ai and Cowork, not buried behind a developer console. That detail matters more than it looks like it should, and I'll come back to it.
A third lever, prompt caching, does more for a real bill than either of the first two, and most launch-day coverage skips it. A request that reuses a large block of context Claude already processed, a system prompt, a codebase, a long uploaded document, can pull that block from a cache instead of paying full input-token price on every call. The API writes the block to the cache once, then reads it back on each later call at a steep discount. Opus 5 charges $6.25 per million tokens to write a five-minute cache and $0.50 per million to read from it, against $5 per million for a fresh, uncached read. An agent that loops over the same codebase a hundred times in a session pays that $5 rate once and the $0.50 rate on every pass after. That mechanic is also why Anthropic built mid-conversation tool changes as a beta feature this cycle, not a random add-on: before this release, changing which tools a running agent could call invalidated the entire cache, forcing a full-price rewrite of everything on the next turn. The cache now survives a tool swap, a narrow fix aimed at one specific, expensive failure mode in long-running agents.
Anthropic tried a version of this "same price, better efficiency" pitch once already this year. It did not hold up under independent measurement. That precedent gets its own section below.
Are you enjoying what you are reading so far? Our staff of writers recommends reading AI Agents Are Breaking the Internet, and Nobody's Ready for the Bill after you finish this one.
Traven
What Opus 5 Claims, and What Backs It Up
Opus 5 ships today at $5 per million input tokens and $25 per million output tokens, the same rate card as its predecessor, Opus 4.8, and half of Fable 5's $10 and $50. A new Fast Mode runs 2.5 times faster for double the price. Cache pricing scales the same way: Opus 5 charges $0.50 per million tokens on a cache hit against Fable 5's $1, according to pricing tables published by The Decoder.
The benchmark numbers back the price story more than I expected going in. Frontier-Bench v0.1 drops a model into a real terminal and scores whether it can plan a fix, run shell commands, read the output, and correct course when something breaks, the same loop a working developer runs, not a multiple-choice quiz about code. A 43.3 percent pass rate on tasks like that represents finishing hard, open-ended engineering work end to end with no human stepping in, which is why that number sits well above Fable 5's 33.7 percent, GPT-5.6 Sol's 34.4 percent, and Opus 4.8's 21.1 percent, more than doubling its own predecessor. GDPval-AA v2, a knowledge-work benchmark, scores models the way chess rates players, as an Elo ranking built from head-to-head comparisons rather than a raw percentage correct, and Opus 5 posts 1,861 against Fable 5's 1,747 and Opus 4.8's 1,593. On ARC-AGI-3, a benchmark built to punish memorized patterns rather than reward them, Opus 5 hits 30.2 percent while GPT-5.6 Sol manages 7.8 percent and Opus 4.8 barely clears 1.5 percent, a gap wide enough that Anthropic itself flags it as an outlier worth independent scrutiny.
Opus 5 doesn't win everywhere, and the losses are specific rather than hand-waved. On DeepSWE v1.1, an agentic coding benchmark, GPT-5.6 Sol leads at 72.7 percent, Fable 5 follows at 69.7 percent, and Opus 5 comes in at 68.8 percent. Anthropic states outright that it chose not to train Opus 5 on cybersecurity tasks, the same choice it made with Opus 4.8, and the model trails Mythos 5 on that front as a result. I'll come back to what that trade-off buys.
One chart in Anthropic's own release argues against maxing out every dial on reflex. On Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index, Opus 5 scores a hair worse at the max effort setting than at the second-highest setting, xhigh, despite max costing more. Spending more tokens on the same task at the top of the dial made the output worse on two of Anthropic's own evaluations. That's not a criticism I'm layering on from outside. It's in the launch material.
The Effort Dial, In Code
The effort field isn't new to Opus 5. It first shipped with Opus 4.5 and has carried across every Opus and Sonnet release since, with the same call shape. A developer named FrankCuiCN filed a bug against it on the Opus 4.6 build in February, and the reproduction case is a clean, real demonstration of how the parameter behaves in a live API call:
# GitHub: https://github.com/anthropics/anthropic-sdk-python/issues/1162
# Author: FrankCuiCN
#
# All credit goes to the above listed person. We are using
# this source code or snippet to provide a real-world example
# for educational purposes.
import anthropic
client = anthropic.Anthropic()
with client.messages.stream(
model="claude-opus-4-6",
max_tokens=16000,
thinking={"type": "adaptive"},
output_config={"effort": "max"},
messages=[{"role": "user", "content": "What is the greatest common divisor of 1071 and 462?"}],
) as stream:
thinking_started = False
response_started = False
for event in stream:
if event.type == "content_block_start":
print(f"\nStarting {event.content_block.type} block...")
# Reset flags for each new block
thinking_started = False
response_started = False
elif event.type == "content_block_delta":
if event.delta.type == "thinking_delta":
if not thinking_started:
print("Thinking: ", end="", flush=True)
thinking_started = True
print(event.delta.thinking, end="", flush=True)
elif event.delta.type == "text_delta":
if not response_started:
print("Response: ", end="", flush=True)
response_started = True
print(event.delta.text, end="", flush=True)
elif event.type == "content_block_stop":
print("\nBlock complete.")
Swap claude-opus-4-6 for claude-opus-5 and this code runs unchanged against the new model. That portability is the point of effort as a design choice: it's a stable contract across the model family, not a one-off flag Anthropic bolted onto a single release.
Two fields carry the whole story. thinking={"type": "adaptive"} hands the model the choice of whether to reason step by step before answering at all, rather than forcing a fixed reasoning budget on every request regardless of difficulty. output_config={"effort": "max"} then tells the model how far to push that reasoning once it decides to use it. The two settings answer different questions. One decides whether the model thinks out loud. The other decides how hard it thinks once it starts. Fable 5 skips the first question and runs with adaptive thinking on for every request, a design choice that fits a model built to chase the hardest tasks regardless of cost. Opus 5 leaves that decision open, which is the more meaningful design signal in this launch than any individual benchmark score: Anthropic built its cheaper, everyday model to spend tokens on purpose rather than by default.
The streaming loop matters as much as the request parameters. Claude can emit a thinking_delta block, a text_delta block, or, per the bug FrankCuiCN reported, an extra empty text block before either of those, depending on the model version. Anyone building a production integration needs to branch on event.delta.type rather than assuming a fixed block order, because that order isn't guaranteed to stay stable across model updates. This is the kind of contract-versus-implementation detail that benchmark charts never mention and that breaks a pipeline in production.
The Cost Math, Worked Out
Run the actual arithmetic instead of trusting the "half the price" headline on faith. A task that burns 5,000 input tokens and 3,000 output tokens on Opus 5 costs (5,000 / 1,000,000 × $5) + (3,000 / 1,000,000 × $25), or $0.025 plus $0.075, for a total of $0.10. The same token counts on Fable 5, at $10 and $50 per million, land at $0.20. Twice the price, as advertised, assuming both models need the same number of tokens to finish the job.
That assumption is where the story gets complicated, and where the rest of this piece lives.
Anthropic's Track Record Problem
Anthropic ran this exact play before, and the receipts from that attempt are the reason I'm not taking "half the price" at face value this time.
Opus 4.7 shipped in April 2026 carrying the identical $5 and $25 sticker as Opus 4.6. Anthropic's own migration guide disclosed a new tokenizer that could produce 1.0 to 1.35 times more tokens for the same input. Independent numbers came in worse than that disclosed range. Developer Simon Willison ran the Opus 4.7 system prompt through Anthropic's token-counting endpoint and found it consumed 1.46 times the tokens of Opus 4.6 on the same text. A community dataset spanning 483 submissions measured a 37.4 percent jump in both tokens and cost switching from 4.6 to 4.7. OpenRouter analyzed more than a million real requests from users who migrated between the two models and found actual billed cost rose 12 to 27 percent for prompts above 2,000 tokens, once they factored in cache absorption. One hands-on benchmark comparing identical coding tasks measured a 3.6x cost increase for the same pass rate, driven by Opus 4.7 taking more agent turns and revising its own work more than 4.6 did to reach the same result.
None of those four measurements come from Anthropic. All four point the same direction: a rate card that stayed flat while the real bill for identical work went up, by a wide margin in at least one case. "Pricing unchanged" and "cost unchanged" turned out to be different claims wearing the same sentence.
Opus 5 could turn out better on this front. The effort dial is a more direct lever than a tokenizer swap buried in a migration note, and Anthropic is now publishing effort-level benchmark curves instead of a single top-line score, a more honest way to present the trade-off than Opus 4.7 got. That possibility is a hypothesis, not a conclusion, and the correct response to a vendor's cost claim with this specific history is to measure your own workload rather than trust the headline number. Anthropic's /v1/messages/count_tokens endpoint is free to call and returns exact counts without running inference. Run your real prompts through it on both models before migrating a production pipeline, the same way you'd benchmark a database index change before shipping it to prod instead of trusting the vendor's changelog.
The Free Tier Never Got the Memo
None of the efficiency story above touches the free tier, and that's worth stating outright because Anthropic's framing makes it easy to assume otherwise. Claude's free plan defaults to Sonnet 5, not Opus. Opus 5 is the new default only on the $100-and-up Max plan, and it's available as an option, not the default, on the $20 Pro plan. A free user hitting a wall after two prompts was never running the model this launch is about.
The mechanism behind that wall is a rolling usage window, not a hard message count and not a fixed daily reset. A quota that resets at midnight gives everyone back their full budget at the same clock time regardless of when they started using it. A rolling window runs on a different principle: it starts counting from your first message and refills over the following five hours, whatever time of day that happens to land on. Anthropic doesn't publish an exact message figure, and estimates from usage trackers land somewhere around 15 to 40 short messages inside that window before the app cuts you off, with a weekly cap layered on top. Length and complexity count against you faster than raw message count does, which is why two dense prompts can burn through a budget that would stretch across a dozen short ones, and why spreading heavy work across the day beats front-loading it into one long conversation.
The timing here is not neutral either. Anthropic cut Fable 5 limits on Max and Team Premium plans to half their previous size starting July 20, four days before this launch, pushing Pro and Team Standard users toward metered API pricing with a one-time $100 credit as a cushion. A company tightening the consumer side of the business right before shipping a model whose entire marketing angle is generosity with tokens is a pattern worth noticing, not an accusation. Both things are true about the same week.
The one lever a free or Pro user has is the effort dial itself, sitting in the Claude.ai interface next to the model picker. Dropping a conversation from the default effort level to medium or low doesn't change what the model knows or its context window, only how hard it works per response, which stretches a token-limited window further. It's a real fix for a narrow slice of the problem, not the whole answer. For anyone whose usage outruns any hosted tier's patience week after week, no matter the provider, running a model on your own hardware removes the rate limit, at the cost of frontier-model capability on the hardest tasks.
The Cybersecurity Carve-Out
Opus 5's cyber safeguards trigger about 85 percent less often than Fable 5's, according to Anthropic, and the model can now research vulnerabilities in source code that Fable 5's classifiers used to block. Binary-based vulnerability scanning, penetration testing, and exploit generation stay off-limits, and any request that trips those classifiers falls back to Opus 4.8 on its own instead of returning a bare refusal.
Anthropic's own automated behavioral audit found Opus 5 comes close to Mythos 5 at finding vulnerabilities in code, while remaining far behind at turning those findings into working exploits. That gap is by design: Anthropic chose not to train Opus 5 on offensive cyber tasks at all. CoderOasis covered what happens when that gap closes instead of holding: Anthropic's Mythos-class models found 271 vulnerabilities in Firefox, ten of them full control-flow hijacks, a scale no prior model had reached. Opus 5 sits on the safer side of that line on purpose, and the fallback behavior to Opus 4.8 is the mechanism keeping it there when a request pushes too close.
Every Claude release this year has shipped a version of this trade-off. The model tier that finds and exploits vulnerabilities at the highest level, Mythos, stays restricted to vetted partners. The model tier everyone gets, Opus, gets the capability boost without the sharpest edge. Opus 5 continues that split rather than changing it.
Anthropic shipped a second beta feature alongside Opus 5 and the cache-preserving tool swap covered above: automatic fallbacks. A blocked request now routes to another model by default instead of returning a bare error, trading a small capability step-down for uptime the user never notices. Combined with the Opus 4.8 fallback already built into the cyber and biology classifiers, a production integration can treat a safety refusal as a routing decision instead of a support ticket.
Four models in under two months, Mythos 5, Fable 5, Sonnet 5, and now Opus 5, is not a company shipping for the sake of shipping. OpenAI landed GPT-5.6 on July 9 across Sol, Terra, and Luna variants, marketing the same restraint story Anthropic is telling today, and open-weight labs like Moonshot AI's Kimi K3 keep closing the capability gap at a fraction of the infrastructure cost of either company. Two of the three largest model vendors spending the same month selling discipline instead of raw benchmark records is a market signal about what buyers are asking for, not a scheduling coincidence. Opus 5's entire positioning, efficient, cheap, tuned for daily use rather than benchmark ceilings, reads like a direct response to that pressure rather than a decision made on its own timeline.
If you're deciding whether to move a workload to Opus 5, the benchmark charts are a reasonable starting point and a bad place to stop. Pull your own traffic through the token counter, run it against both models, and let your actual bill settle the argument the rate card can't.