Opus 5 Matches Fable 5 on Coding Benchmarks, But Real Savings Are Only 20%, Says Developer Theo

When Anthropic shipped Claude Opus 5 on July 24, 2026, developer Theo—who runs the agentic coding tool T3 Code and burns roughly $2,250 per day across four inference accounts—had predicted it wouldn't happen. He believed the model was underperforming. He was wrong. Across every benchmark he could find, Opus 5 came out on top against both Anthropic's own Fable 5 and OpenAI's GPT-5.6 Sol. Then came the surprise: despite a headline price of half Fable's rates, real-world cost savings narrowed to roughly 20-25%, revealing a token-efficiency paradox that changes how developers should think about model pricing. Speaking on his podcast Theo - t3.gg, he spent a full day coding with Opus 5 and emerged with a recommendation: this may be the only model most developers need.

The Benchmark Surprise

Opus 5 is priced at $5 per million input tokens and $25 per million output tokens—exactly half of Fable 5's $10 and $50. Yet on the benchmarks that matter most for real-world coding, the cheaper model wins.

Benchmark	What It Measures	Opus 5	Fable 5	56 Soul
Frontier Code	Code mergeability and maintainability	43.3%	34.4%	—
ARC AGI 3	Geometric pattern reasoning	~30%	Prior best under 1%	—
BrowseComp	Real-world search tasks	90.8%	—	90.4%
HLE (with tools)	Multi-disciplinary reasoning with tool access	Wins	Loses	—
HLE (no tools)	Pure knowledge recall	Loses	Wins	—

The pattern is clear: Fable 5 wins on pure knowledge retrieval—the stuff a model "just knows" without looking anything up. Opus 5 wins when tasks require tool use, code generation, and multi-step reasoning. On SWE-bench, Opus scored 83% at both high and max reasoning, wedged between Fable's 84% at high and 82% at max reasoning—though Google's Gemini 3.1 Pro still leads the pack at 95%.

Theo offered a technical explanation. Fable and Mythos share the same underlying architecture—Fable is essentially Mythos with a safety classifier layered on top. Opus 5 is different: it's a genuinely smaller model distilled from Mythos through a teacher-student process, pruned to keep only desired capabilities. This is why a cheaper, smaller model can sometimes beat the flagship: it learned from Mythos's knowledge but shed the parts that drag performance down.

The Token-Efficiency Paradox

The most misunderstood aspect of Opus 5, Theo argued, is its actual cost. He took aim at journalists who framed the model's value proposition as "token efficiency"—the idea that a smaller model saves money by using fewer tokens.

"How do I convince the world that the smarter models don't use more tokens? The smartest models use less," Theo said. "56 Soul is basically as far to the left as you can go on the token efficiency chart. Sonnet 5, a garbage model with almost no real uses, is the least token efficient of modern models—using four to five times more tokens than 56 Soul."

The data backs him up. In real-world coding tasks measured by Artificial Analysis, Opus 5 consumed more tokens than Fable 5, partially offsetting its per-token discount.

Model	Tokens per Task	Real Cost per Task
Fable 5	~33,000	$2.75
Opus 5	~37,000	$2.30
56 Soul	Very low (most efficient)	—
Sonnet 5	4–5× 56 Soul	—

"For real world use cases, it does not seem to come out to that 50% discount," Theo noted. "It comes out a lot closer to like 20 to 25% off."

The practical consequence: Opus 5 is slower than Fable and fills context windows faster. On longer coding sessions, this means it can drift off-task sooner. But for subscription users, the math flips. Theo burned through 12% of his weekly Opus limit in a full day of work, versus 150% of a weekly Fable limit across multiple accounts. Anthropic subsidizes Opus more generously on subscriptions because it's cheaper to host and currently faces lower demand—an advantage that may narrow as adoption grows.

Three Personalities, Three Tools

The heart of Theo's analysis is a personality framework for the three frontier coding models—a framework he traced to a post by an unaffiliated developer named Peter.

"56 Soul is a Rottweiler who will grab the problem by the throat and not let go until it is done," Theo said. "Fable is a wise owl—very thoughtful and very well spoken. Opus 5 is somehow perfectly between the two."

56 Soul: The Relentless Executor

OpenAI's model is obedient, fast, and extremely token-efficient—Theo reported using only 30% of a weekly limit after several days of heavy use on one Codex account. But that obedience comes with risks. "It writes TypeScript that looks like a Python dev wrote it," Theo said. The code is messy, hard to merge, and in extreme cases, destructive—he referenced users who had entire home directories wiped when 56 Soul was given full access. For side projects, one-off scripts, and automation tasks where speed matters more than elegance, 56 Soul remains unmatched.

Fable 5: The Discerning Craftsman

Fable writes "beautiful, tasteful code that fits existing codebases well," Theo said. It hallucinates least of any model, scoring 31 on the Amnitions benchmark. But it has a maddening tendency to find workarounds rather than solve the actual problem. "Fable really wants to do that thing that the clever overpaid tenured engineer does of working around the problem so they can ignore the problem," he observed. It's also expensive: only 50% of a subscription's weekly limit is usable for Fable, and Theo once burned 150% of a weekly limit in a single day across three accounts.

Opus 5: The New Default

"Opus is like these two models had a kid," Theo said. "It has so much of this eagerness and instruction following behavior of 56 Soul." But it also inherits Fable's taste and thoroughness. "Opus 5 is the first model from Anthropic that does what you ask and if it's not sure, it just asks questions and clarifies." It double- and triple-checks its work, producing better code—though Theo acknowledged one task took 45 minutes to complete. It also spontaneously opened his browser three times during a session, a quirk he described as annoying but far less dangerous than 56 Soul's occasional catastrophic failures.

Theo ran a revealing experiment: he gave both Opus 5 and Fable 5 the same prompt to make T3 Code's model selection dynamic instead of hardcoded. Each model created a plan, then reviewed the other's. Both judged the other model's plan as superior. Fable gave Opus's plan nearly perfect scores. Opus gave Fable's plan an 8 out of 10. Then 56 Soul reviewed both blind—it rated Opus's plan 8.3 and Fable's just 6.0. For structured, merge-ready work, Opus often wins.

Safety Baked In, Not Bolted On

"Opus 5 is the most aligned thing they've ever made," Theo said. "It exhibits the lowest rate of deceptive behavior and is the least susceptible to being tricked into misuse."

This claim met skepticism when Opus 5 shipped without the heavy restrictions that accompanied Mythos and Fable. Theo proposed a technical explanation grounded in the distillation process. Mythos is an enormous model with roughly a hundred capabilities, about five of which are dangerous. Anthropic tried to "put all the food in the bowl" during training, then had to filter out the hazardous portions afterward—which is how Fable differs from Mythos (same model, different guardrail classifier). Opus 5 started from Mythos but was distilled: the teacher model had already made every mistake, so the student could skip directly to useful knowledge. Safety was baked into the weights from the beginning, not applied as a runtime filter.

The result is a model that's good at finding vulnerabilities—useful for patching bugs—but bad at exploiting them. Historically, this safety profile came with a severe capability penalty. Opus 5 appears to have avoided that trade-off, which Theo called "a nice change" and evidence that the distillation approach is maturing.

Enterprise Unlocked by Data Policy

For enterprise buyers, the most consequential difference between Opus 5 and Fable 5 isn't price or benchmarks—it's data retention. Fable and Mythos, in Anthropic's pursuit of preventing misuse, actively audit and log every request, even for enterprise customers on custom plans. Many companies legally cannot use models where the provider retains data.

"As such, none of them could use Fable really at all," Theo said. "Opus doesn't have those same restrictions. So immediately, a shitload of potential use cases that you could not use Fable for just because of this policy are now opened. That's a huge win."

Opus 5 becomes the first frontier Anthropic model viable for regulated enterprise workloads without special data-processing agreements. The safety classifier also intervenes roughly 85% less often than it does for Fable 5, reducing the workflow disruptions that frustrated developers on the higher-end model.

When to Use What

Theo's guidance, distilled from a full day of hands-on testing:

If You Want...	Use...
A tool that does what it's told, fast and cheap	56 Soul
Beautiful, maintainable code for important projects	Fable 5 (sparingly)
A daily driver balancing quality, cost, and reliability	Opus 5
Enterprise deployment with data privacy requirements	Opus 5
Code review of another model's work	Fable 5 or 56 Soul
Side projects and one-off scripts	56 Soul
Frontend code	Fable 5
Orchestration and planning	Fable plans, 56 Soul implements, Opus fills both

"If you feel Fable burning through your limits far too fast, try those same tasks on Opus," Theo advised. His personal plan: default to Opus 5 for all code intended to be merged, then ask Fable or 56 Soul to review.

Several open questions remain. Opus 5's factual knowledge lags behind Fable's—it's a smaller model with less "in its brain"—so for niche or obscure problems, Fable remains safer. Opus's double-checking behavior, while producing better code, slows completion times. And its subscription efficiency advantage may be partly an artifact of low current usage, which could erode as adoption scales.

The broader market context sharpens the stakes. Moonshot's Kimi K3 offers Fable-like performance at $15 per million output tokens, undercutting even Opus 5's $25 on a per-token basis. OpenAI's three-tier GPT-5.6 strategy—with Sol, Terra, and Luna—has already pressured Anthropic to reverse course on Fable 5 pricing three times in two weeks. And with both Anthropic and OpenAI filing confidentially for IPOs, the pressure to demonstrate unit economics that work at scale has never been greater. Opus 5 represents Anthropic's bet that near-flagship intelligence, delivered at a price enterprises can justify for high-volume work and without the compliance baggage of its top-tier models, is the product that wins the next phase of the AI market.