Marcus Vale: Ben, hey — you catch any of the Anthropic coverage this week, or did it just wash over you like everything else?
Ben Okonkwo: Hm, actually — I was reading the launch post on the train Thursday and I genuinely stopped scrolling. Not because of the model. Because of the price. It's five dollars per million input tokens. Same as Opus 4.8. And my first thought was, wait, that's... that's too clean.
Marcus Vale: Same sticker. Smarter model. That's the headline Anthropic wants.
Ben Okonkwo: Right — but the part that doesn't fit is the effort dial. Anthropic ships Claude Opus 5 on July 24th, same five-dollar, twenty-five-dollar token pricing as before, and then quietly hands you a knob that controls how hard the model thinks. Turn it up, you get near-Fable 5 performance. Turn it down, you save tokens. So the price isn't actually fixed. It's... contingent.
Marcus Vale: Here's the deal — think about a gym membership. Flat monthly fee. Sounds great. But the fee covers a treadmill, not a personal trainer. The trainer costs extra, only nobody tells you what a session runs until the bill arrives. The effort dial is the trainer. That's the actual pricing structure.
Ben Okonkwo: Okay, but — I want to push back on that framing a little. The sticker price genuinely is the same. Five and twenty-five per million tokens, identical to Opus 4.8. You're calling the feature a trick when it might just be... a feature. OpenAI has compute controls too.
Marcus Vale: The feature creates invisible cost variance. That's my point. The CFO approves the API spend based on the token rate — which is fixed — but the real cost per solved problem is variable and nobody's disclosing the conversion. That's not the same as a feature.
Ben Okonkwo: So the question we're actually trying to work out is whether Anthropic's efficiency story holds — or whether Claude Opus 5's pricing is structured to obscure what near-Fable 5 performance actually costs you.
Marcus Vale: And the proof that it obscures cost is sitting right there in the Frontier-Bench numbers. Opus 5 hits 43.3% on Frontier-Bench v0.1. Opus 4.8 was at 18.7%. Same token price. That's not incremental — that's more than double the capability at zero sticker increase. So which tasks burned the extra compute to get there? Nobody's publishing that conversion.
Ben Okonkwo: Wait — I'd actually flip that. The 43.3% versus 18.7% is my evidence, not yours. That jump is real. That's a measurable capability gain at the same price point. The effort dial isn't hiding a cost — it's what made that jump possible without raising the rate.
Marcus Vale: No, I don't buy that.
Ben Okonkwo: And the mid-task model-switching feature — that's actually meaningful cost management. You start a task on Opus 5, it gets gnarly, you switch to Fable mid-run. Or it turns out to be simple, you dial down. OpenAI had compute controls first, Anthropic is mirroring it — but the mid-task switching is a layer beyond that. That's... actually giving the buyer a lever.
Marcus Vale: A lever nobody can model in advance. Look — Amazon Bedrock hosts Opus 5 with zero data retention by default. That's the enterprise pitch. But ZDR default plus a variable effort setting means a CFO approving Bedrock API spend is budgeting against... what, exactly? A token rate that's fixed and a task cost that isn't.
Ben Okonkwo: Hm. The ZDR point is — okay, I hadn't connected those two things.
Marcus Vale: Zero data retention means no usage telemetry leaving the enterprise perimeter, right? Which sounds great for compliance. But it also means the buyer has even less visibility into what the effort dial actually consumed per task. The opacity is structural, not accidental.
Ben Okonkwo: That's — okay, that's the load-bearing assumption I'd want to test. Because the mid-task switching and the dial are the same architecture. If performance scales linearly with effort, the buyer can at least calibrate. If it cliffs — if you only get the 43.3% at maximum effort — then you're right, the benchmark number is only reachable at a cost nobody disclosed.
Marcus Vale: And Anthropic isn't publishing the cliff point. That's the bet I'd make.
Ben Okonkwo: The cliff point is real — but I think you're burying the more interesting thing underneath it. Opus 5 scores higher than Fable 5 on coding and knowledge work. Fable stays flagship. Anthropic explicitly says Opus 5 is not state-of-the-art for risky dual-use capabilities. That's not a benchmark caveat. That's the company telling you the benchmarks don't measure what they're actually optimizing for.
Marcus Vale: So Fable's position is a safety designation, not a capability designation.
Ben Okonkwo: Which is — okay, it's actually honest. But it's also saying the whole efficiency narrative Anthropic built around Opus 5 rests on a category of tasks where the security ceiling doesn't apply.
Marcus Vale: And three days after Opus 5 launches, Microsoft announces MAI-Cyber-1-Flash on July 27th, claiming it beats Claude Mythos 5 on CyberGym at half the cost. That's not a coincidence. That's a strike directly on the gap Opus 5 can't close.
Ben Okonkwo: Wait — half the cost of Mythos 5?
Marcus Vale: Half. And SecurityWeek already reported Opus 5 nears but does not match Mythos 5 on bug-finding, falls short on exploit tasks entirely. So Anthropic's security moat is Mythos — and Mythos just got flanked. The high-margin enterprise security contracts go where the exploit capability is.
Ben Okonkwo: Right, but — I mean, that's the mechanism I'd actually want to stress-test. CyberGym is one benchmark. Does MAI-Cyber-1-Flash hold that edge on real adversarial tasks, or is this... another benchmark number that cliffs under production conditions?
Marcus Vale: Fair. But Boris Cherny is out here telling users to stop micromanaging Claude Code and trust the model with higher-level agentic tasks — right as the security model that underpins that trust just got outbid on the exact capability it owns.
Ben Okonkwo: That's the tension, yeah. The product philosophy and the competitive position are pointing in opposite directions.
Marcus Vale: And then there's the fourth Claude 5 model in under two months — Fable, Mythos, Sonnet, now Opus — and nobody's asking how deep the safety evaluation goes per release. That question gets a lot harder when we get to what happened with the shared chats.
Ben Okonkwo: The shared chats thing is — okay, that's the incident. Claude shared chats were indexed by Google, surfaced on Reddit. During an active enterprise sales push. That's not a theoretical trust wound. That's a specific, dated failure.
Marcus Vale: And Anthropic's entire enterprise moat is built on the safety-first brand. That's the thing. The closed-source positioning, no open weights — the argument is that staying closed is how they protect data governance. And then the data leaks anyway.
Ben Okonkwo: I'll — actually, I'll concede that point directly. The privacy incident is a real wound. You can't sell Amazon Bedrock with zero data retention as the trust argument and then have shared chats show up on Reddit in the same news cycle.
Ben Okonkwo: But — and this is where I'm holding the line — data governance and a privacy incident aren't the same category. Enterprise compliance departments aren't asking whether Anthropic has perfect execution. They're asking whether Anthropic has contractual control. Open-weight alternatives can't give them that. The Bedrock ZDR default is a liability instrument, not a feature.
Marcus Vale: What's the liability instrument worth when Dario Amodei can't actually tell you what architectural property open-weighting would compromise? That's the unfalsifiable part. Anthropic never specifies what breaks if they release the weights.
Ben Okonkwo: That's — no, that's fair. The safety justification for staying closed is genuinely unfalsifiable without knowing the specific property. I can't defend that gap. What I can say is that enterprise buyers aren't waiting for Dario to answer it — their legal teams already chose the contractual model. That's not nothing.
Marcus Vale: Until an open-weight competitor matures enough to offer the same liability wrapper. Then the moat leaks the same way the shared chats did.
Ben Okonkwo: I think where I actually land is — uncomfortable, honestly. Anthropic built the safety brand, and now the competitive pressure is forcing them onto price. Four Claude 5 models in under two months. That's not a safety-first cadence. That's a company being pushed.
Marcus Vale: Yeah. And the confession is in the pricing structure itself. Same sticker. Variable real cost. The effort dial doesn't hide the cost. It just moves it somewhere the CFO can't see yet.
Ben Okonkwo: Or it's the only rational move when your competitors are OpenAI and Google and Microsoft is three days behind you with a cybersecurity model undercutting Mythos 5. I mean — what's the alternative? Stay principled and lose the slot?
Marcus Vale: That's the bet Anthropic is making. We'll find out if it holds.
Ben Okonkwo: Good talk. Genuinely — I walked in thinking the effort dial was just a feature. I'm leaving less sure of that.
