Idea #1
Cost per task — not price per token — is the only metric that tells you what you're actually paying for AI work. Price per token is a seductive shortcut because it's printed on the pricing page, but it hides token efficiency differences that can completely erase a cost advantage. If Model A charges half the price per token but consistently uses twice the tokens to complete the same task, the real price is identical. This matters enormously when evaluating competing models, because a cheaper-looking model can be a wash or even more expensive at the task level. The framing also sharpens what to optimize: you want models that solve problems in fewer tokens, not just models with lower unit prices. The concrete example from the source is Kimi K3: it was priced at less than half the cost of Fable 5 per token, yet benchmarks showed it requiring roughly twice the tokens to accomplish the same task, making its effective cost equivalent. Anthropic explicitly leaned into this framing with Opus 5, presenting charts with 'cost per task' on the x-axis and pass rate on the y-axis — showing Opus 5 as both cheaper and higher-quality than GPT-5.6 Soul and Fable 5 on the OS World computer-use benchmark.
[from: What did Anthropic do?! (Opus 5) — YouTube · Matthew Berman · https://youtu.be/tHQ34j8_toI]
Idea #2
Safety constraints on a model do not necessarily degrade its general capability — and in some cases the opposite holds. The conventional assumption is that guardrails come at a cost: you reduce what the model can do in a restricted domain, and bleed-over makes it worse everywhere else. This may be true for crude filtering approaches, but a model trained or fine-tuned with more surgical intent can have targeted capability reductions in a specific domain while improving on nearly every other dimension. This challenges the instinct to treat 'unconstrained' models as capability benchmarks — a model with fewer dangerous capabilities can still be the strongest general performer. The source gives a direct example: Opus 5 has substantially reduced cyber-exploitation capability versus Mythos 5 (which appears to operate with no guardrails on cyber tasks), scoring only 4% on exploitation success versus Mythos's 13% and Opus 4.8's 0%. Yet Opus 5 outperforms Fable 5 — Anthropic's previously top-tier model — on nearly every other benchmark including coding, computer use, and practical reasoning. The speaker flags this as 'crazy' and counterintuitive, noting that typically constraining a model hurts it broadly, but the data shows otherwise here.
[from: What did Anthropic do?! (Opus 5) — YouTube · Matthew Berman · https://youtu.be/tHQ34j8_toI]
Idea #3
The pattern of 'frontier model first, then optimize' is a deliberate product strategy, not an accident — and it means you should not form long-term cost opinions from a model's first release. When a new architectural or training approach is deployed for the first time, the goal is validating capability, not efficiency. Token usage, inference cost, and price-per-task are all secondary. Once the capability envelope is established, subsequent model iterations can apply the same techniques with far better efficiency. If you locked in a negative view of a model family because the first version was expensive, you may be pricing yourself out of a better deal by the third iteration. The source illustrates this with Fable 5 and Opus 5: Fable 5 drew criticism for high token usage and cost. The speaker had predicted publicly that Anthropic would 'optimize the hell out of it,' and his reading of Opus 5's launch is that Fable served as the proving ground — Anthropic used it to push the capability frontier, then applied those learnings to Opus 5, delivering superior benchmark scores at half the price. Opus 5 is priced at $5 input / $25 output per million tokens, matching Opus 4.8 and roughly half of Fable 5.
[from: What did Anthropic do?! (Opus 5) — YouTube · Matthew Berman · https://youtu.be/tHQ34j8_toI]
Idea #4
At sufficiently high thinking levels, model performance can plateau or even decline while cost increases — meaning maximum thinking is not always the right setting. This is a practical calibration issue for anyone building agentic systems or using extended thinking features: more compute is not monotonically better, and the optimal setting is empirical, not 'max by default.' The cost implication compounds: you can pay more and get worse results if you haven't tested which thinking level actually performs best for your task type. The source shows this on Frontier Bench data for Opus 5: moving from a mid-range thinking level to the highest thinking level produced a lower benchmark score at a higher price. By contrast, a step just above mid-range gave a roughly 5% score improvement over GPT-5.6 Soul's maximum for only about $2 more per task. The takeaway the speaker drew is that practitioners need to treat thinking level as a tunable hyperparameter tested against their specific tasks — not as a slider to leave at maximum.
[from: What did Anthropic do?! (Opus 5) — YouTube · Matthew Berman · https://youtu.be/tHQ34j8_toI]
Source: What did Anthropic do?! (Opus 5) (Matthew Berman)
Text summarized and optimized using Anthropic’s models and reviewed by a human.