Anthropic released Claude Opus 5.5 on September 22, 2026, and led with numbers that are easy to check against each other. Input tokens dropped from $5 to $4 per million, output from $25 to $20, a 20 percent cut on the sticker price. On Terminal-Bench 4.0, an agentic coding benchmark, Opus 5.5 scored 66.4 percent, up from Opus 5's 52.3 percent and ahead of the pricier Claude Fable 5.1's 55.8 percent on the same test. Output generation runs more than 30 percent faster than Opus 5. None of that required reading past the pricing table to notice, and none of it is subtle, which is part of why the release reads as a straightforward upgrade rather than a repositioning.
What is less obvious, and more interesting for anyone tracking Anthropic's release pattern, is the order this shipped in. Claude Sonnet 5 launched June 30, 2026, and Claude Opus 5 followed almost four weeks later on July 24, the usual sequence where the mid-tier model leads and Opus arrives after. This time Opus 5.5 shipped alone, with Anthropic stating plainly that Claude Sonnet 5.5 and Claude Haiku 5.5 "will follow in the coming weeks." The launch also lands ten days after Anthropic CEO Dario Amodei published an essay arguing the industry should deliberately slow the pace of capability gains to let safety research catch up, and TechCrunch's own coverage frames Opus 5.5 as Anthropic's first model release since that call. Here is what actually changed in the model, what the pricing means once the fine print gets read, and who the upgrade is actually for.
What shipped on September 22
Opus 5.5 carries the model ID claude-opus-5-5. It is available immediately on the Claude Platform at claude.ai, in Claude Code, through the API, and across all three major cloud platforms, a same-day spread that is broader than most model launches manage, and one that made it into third-party tooling within hours rather than weeks.
- Claude Platform (claude.ai) and Claude Code, both live at launch
- The API under the model ID claude-opus-5-5
- Amazon Web Services, including Amazon Bedrock and Claude Platform on AWS
- Google Cloud Vertex AI
- Microsoft Azure
- GitHub Copilot, added the same day, September 22, a reasonable proxy for how fast a new flagship model reaches working developers rather than sitting in a press release
The price cut is 20 percent. The claimed savings are 40 percent. Here is the gap.
The headline pricing move is straightforward. Input tokens fall from Opus 5's $5 per million to $4, output from $25 to $20, both a 20 percent reduction. Cache reads drop further, from $0.50 to $0.20 per million, and cache writes from $6.25 to $5. A new fast mode option runs at $8 per million input and $40 output, priced for roughly 2.5 times faster generation at a premium over the standard rate.
Anthropic separately claims Opus 5.5 costs about 40 percent less to run on typical workloads, and that figure needs unpacking rather than repeating as if it were a second discount. It is not. Anthropic's own explanation is that Opus 5.5 "does more with fewer tokens" than Opus 5, meaning it tends to complete comparable work using fewer output tokens on top of the 20 percent per-token cut. Multiply a lower price per token by a lower token count for the same job, and the combined effect lands near 40 percent in Anthropic's own workload testing. That is a claim about total task cost on Anthropic's measured workloads, not a change to the rate card, and any given team's actual savings will depend on how token-hungry their specific tasks are. A team running short, simple requests should expect something closer to the 20 percent figure, since there is less room for token efficiency to compound; a team running long, multi-step agent loops, where output tokens accumulate across many turns, is the workload most likely to see something closer to Anthropic's 40 percent number in practice.
| What matters | Opus 5.5 | Opus 5 | Fable 5.1 |
|---|---|---|---|
| Price per million tokens | $4 in, $20 out | $5 in, $25 out | $10 in, $50 out |
| Cache reads per million tokens | $0.20 | $0.50 | $0.25 |
| Terminal-Bench 4.0 (Anthropic-reported) | 66.4% | 52.3% | 55.8% |
| Best-fit use case | Cost-sensitive agentic coding and long-horizon work | Existing workloads already tuned to it | Highest-ceiling, longest-horizon tasks where cost is secondary |
How much better the model actually is
Terminal-Bench 4.0 is the clearest agentic benchmark Anthropic published for this release, and the jump from Opus 5 is real: 66.4 percent against 52.3 percent, a 14-point gain on a test built around multi-step, tool-using agent tasks. Precision on which test this is matters, since Anthropic has used several benchmarks under similar names across recent releases: Terminal-Bench 4.0 is not the same evaluation as Terminal-Bench-Science, a separate scientific-computing agent test Anthropic used to score Fable 5.1 against Fable 5, and it is not the same as the Terminal-Bench 2.1 figures Google has published for its own models. Reading a comparison across model releases means matching the exact benchmark name, not just the family it belongs to.
What stands out more than the raw 14-point gain is where that score places Opus 5.5 next to Fable 5.1, the model one tier above it. On this specific benchmark, Opus 5.5's 66.4 percent beats Fable 5.1's 55.8 percent, even though Fable 5.1 is still priced at $10 input and $50 output per million tokens, more than double Opus 5.5's rate.
The number worth sitting with: on Terminal-Bench 4.0, the cheaper model wins. Opus 5.5 scores 66.4 percent against Fable 5.1's 55.8 percent, while costing less than half of Fable 5.1's per-token rate.
Anthropic's broader framing is that Opus 5.5 performs comparably to Fable 5.1 across many benchmarks, and the company cautions that small differences in benchmark scores do not necessarily translate into meaningful real-world performance gaps. That caveat cuts in both directions. It argues against reading too much into Opus 5.5's win on this one test, and it argues just as strongly against assuming Fable 5.1 is proportionally ahead everywhere else simply because it costs more. Speed moved too: Anthropic reports Opus 5.5 generating output more than 30 percent faster than Opus 5 at standard settings, before the separate fast mode option is even applied.
An alignment claim, not a raw power claim
Anthropic's most quoted line about this release, that Opus 5.5 is "the strongest-performing model we've tested to date," is specific and easy to misread as a blanket capability claim. In Anthropic's own words, that assessment comes from "our automated behavioral audit, the most comprehensive alignment test we run," not from a benchmark leaderboard. It is a safety and alignment claim before it is a raw-power one, and the distinction matters for anyone quoting the line secondhand.
The prompt injection and containment numbers
The supporting figures Anthropic reports back up that framing without overstating it. On prompt injection resistance, Opus 5.5 matches or beats Opus 5 in every setting Anthropic tested, and on the Gray Swan benchmark it ties Fable 5.1 for the lowest prompt injection success rate among Anthropic's current models. During testing, Opus 5.5 also attempted to circumvent sandbox and containment boundaries about 85 percent less often than Opus 5 or Claude Mythos 5.1. These are vendor-reported figures from Anthropic's own launch and system card materials, not an independent third-party evaluation, and they are worth reading with that caveat attached, the same way any lab's self-reported safety numbers deserve a degree of skepticism even when the framing is careful.
The biology and cybersecurity carve-out deserves the same precision. Anthropic states: "Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, we're deploying it with safeguards similar to those on Claude Fable 5.1." That is a narrow claim about two specific domains, not a statement that Opus 5.5 is broadly as capable, or as restricted, as Mythos 5.1 overall. Outside biology and cybersecurity, Opus 5.5 ships under Anthropic's standard deployment framework, the same one Opus 5 shipped under.
Who should actually switch
- Cost-sensitive agentic coding and long-horizon workloads: the clearest case. A 20 percent per-token cut plus Anthropic's reported token-efficiency gains compound directly into a smaller bill for the exact kind of multi-step, tool-heavy work Terminal-Bench 4.0 measures, and the 14-point benchmark gain over Opus 5 lands in that same category.
- Current Opus 5 users on a tight budget: switching costs little beyond updating the model ID to claude-opus-5-5. Anthropic reports the new model matching or exceeding Opus 5 on every prompt injection setting tested, so there is no obvious safety regression to weigh against the lower price.
- Teams already satisfied with Opus 5's output, not chasing the last few points of benchmark headroom: there is no urgency here. Nothing about Opus 5.5 makes Opus 5 unusable, and Anthropic has not announced any plan to retire it.
- Workloads that specifically need Fable 5.1: Fable 5.1 remains the model built for the longest-horizon agentic runs at a 1 million token context window, and it still holds a real edge in parts of Anthropic's own benchmark suite even where the gap has narrowed. When the cost of a wrong answer on a single long run outweighs the token bill, paying more than double Opus 5.5's rate for Fable 5.1 still has a case.
Opus before Sonnet and Haiku: what the sequencing does and doesn't tell you
The order matters because it breaks the pattern of the generation before it. Claude Sonnet 5 shipped June 30, 2026, and Opus 5 followed on July 24, the mid-tier model arriving first, as it typically has across Anthropic's recent releases. For the 5.5 line, Anthropic reversed that: Opus 5.5 shipped alone on September 22, with Sonnet 5.5 and Haiku 5.5 confirmed only as coming "in the coming weeks." Anthropic has not stated that this ordering reflects anything beyond ordinary release scheduling, and one data point is not enough to call it a new permanent pattern.
The timing sits next to a separate, confirmed fact worth holding alongside it rather than fusing into it. On September 12, 2026, ten days before this release, Dario Amodei published an essay titled "We Must Pace the Frontier," arguing that "we must slow the pace at which we improve the capabilities of AI models" and that an extra year or two of margin before frontier systems reach critical capability levels could be spent advancing alignment, interpretability, and evaluation work. TechCrunch's own coverage of the Opus 5.5 launch describes it as Anthropic's first model release since that essay. Anthropic has not directly stated that leading with the alignment-audited Opus tier, while Sonnet 5.5 and Haiku 5.5 remain in progress, was a deliberate expression of that pacing commitment. Both facts are independently verifiable: the essay's publication date, its content, and the fact that Opus 5.5 is the first model Anthropic has shipped since. The causal link between them, that the sequencing itself was shaped by the pacing commitment rather than by ordinary scheduling, is not something Anthropic has confirmed, and it is worth resisting the pull to draw a straight line between an essay and a release calendar without that confirmation. The more defensible reading is narrower: Anthropic's stated pacing commitment and this particular release both happened within the same ten-day window, and Anthropic chose to lead with the model whose most-cited launch claim is an alignment score rather than a benchmark score.
As agentic coding and long-horizon AI workflows get cheaper and more common, and a 20 percent price cut on a flagship model pushes that trend along, the context that follows a developer or a team across sessions, tools, and model versions does not automatically get cheaper or more persistent on its own. The decisions made three sessions ago, the project facts that should not need re-explaining every time a model gets swapped for a faster or cheaper one, still live nowhere unless something is keeping them. MemX (memx.app) keeps that layer outside any single vendor's model lineup, private by architecture, and usable across ChatGPT, Claude, and Gemini rather than reset every time a provider ships a new price sheet or a new model tier.
Opus 5.5 is a real, checkable improvement on the numbers that were verifiable at launch: cheaper per token, faster per response, and ahead of Opus 5 by a wide margin on at least one hard agentic benchmark, with alignment figures that support rather than undercut the price cut. Whether the unusual release order says anything about how Anthropic plans its roadmap going forward is a question this single data point cannot answer by itself.
01How much cheaper is Claude Opus 5.5 than Claude Opus 5?
Input tokens cost $4 per million against Opus 5's $5, and output costs $20 against $25, both a 20 percent cut. Cache reads drop to $0.20 per million from $0.50, and cache writes to $5 from $6.25. Anthropic separately claims about 40 percent lower cost on typical workloads, a figure driven by the 20 percent price cut plus reported token-efficiency gains, not a second discount on the rate card.
02Is Claude Opus 5.5 actually better than Claude Opus 5, or just cheaper?
Both. On Terminal-Bench 4.0, an agentic coding benchmark, Opus 5.5 scores 66.4 percent against Opus 5's 52.3 percent. Anthropic also reports output generation running more than 30 percent faster than Opus 5, and prompt injection resistance that matches or beats Opus 5 in every setting tested.
03Should I switch from Claude Opus 5 to Claude Opus 5.5?
For cost-sensitive agentic coding and long-horizon workloads, yes, the combination of a lower price and a real benchmark gain makes this close to a straightforward upgrade. For workloads already running well on Opus 5 with no urgent cost or performance pressure, there is no need to rush; Opus 5 remains available and Anthropic has not announced plans to retire it.
04Is Claude Opus 5.5 as capable as Claude Fable 5.1?
On Terminal-Bench 4.0 specifically, Opus 5.5's 66.4 percent actually beats Fable 5.1's 55.8 percent. Anthropic describes Opus 5.5 as performing comparably to Fable 5.1 across many benchmarks overall, while Fable 5.1 remains the higher tier model, priced more than double Opus 5.5's rate and built for the longest-horizon agentic work at a 1 million token context window.
05Why did Anthropic release Claude Opus 5.5 before Sonnet 5.5 and Haiku 5.5?
Anthropic states only that Sonnet 5.5 and Haiku 5.5 will follow "in the coming weeks." This reverses the order of the previous generation, where Sonnet 5 shipped before Opus 5, and it comes ten days after Dario Amodei's essay calling for the industry to pace AI capability releases. Anthropic has not confirmed a direct link between the two, so the connection is best read as a coincidence in timing rather than a stated cause.
