Opus 5.5's claimed 40% cost cut needs a cache-heavy workload to appear

Take the list prices backward. If $4 and $20 per million tokens sit 20 percent under the previous flagship, Opus 5 was $5 and $25 ^[3][1]. A cache read at $0.20 that is 60 percent cheaper than before was $0.50 ^[4][2]. Inside the new price list, a cached input token costs a twentieth of a fresh one ^[3].

Cutting input and output prices by 20 percent gets a workload halfway there ^[6]. The remaining 20 points have to come from the cache line, and the two published cuts bracket the answer: an all-fresh context sees 20 percent, and a context that is almost entirely cache hits approaches 60 ^[7]. Forkast reports one run of each task and does not define the typical workload behind the 40 percent ^[2]. Output speed is up more than 30 percent, which is a latency figure and not a token count ^[2]. Subscription limits on Pro, Max and Team went up by roughly 25 percent ^[16].

Then the port. 9.5 hours against 12 is a 21 percent reduction in wall-clock time ^[5][4], and Forkast calls it a 51 percent improvement in the same sentence that gives both hour counts ^[5]. The 51 has to be measuring something else. The audit comparison is coarser and larger: under three hours against 20-plus, at least an 85 percent cut at the stated bounds ^[6][5].

The benchmark row is 66.4 percent on Terminal Bench 4.0, 54.4 on FrontierCode v1.1, 57.8 on CursorBench 4.0 and 40.0 on AutomationBench ^[7]. HLE with tools is 67.7 percent, 10.5 points above GPT-6 Astra by Forkast's account ^[8]. With tools means the scaffold was part of what was measured, so the score moves when the harness does. A pass rate also averages over retries, and a team repricing a pipeline needs cost per completed task. The OSWorld 2.0 entry appears as an 81.8 percent lead, a phrasing that could describe a score or a margin, and GDPval-AA v2.1 is given as 1846 Elo ^[9].

The timing sits awkwardly beside the company's own framing. Dario Amodei published "We Must Pace the Frontier" on September 12, and the 5.5 family shipped ten days later ^[13][1][8]. Forkast places the launch on the day of the BC AG v. OpenAI filing, three days after reports of an Anthropic IPO and four days after an antitrust suit filed September 18 ^[14][9], and reads the sequence through what it calls the Bessent management-responsibility doctrine, which asks a company to show dominance and responsible management at the same time ^[15]. That reading is the publisher's. The external behavioural audits it points to were run by Frontier Design and METR ^[11].