Claude Opus 5.5

Claude Opus 5.5 is a large language model by Anthropic, and the first model in the Claude 5.5 generation. It was released September 22, 2026, about eight weeks after Opus 5 and three weeks after Mythos 5.1. Anthropic positioned it as performing “at the level of Claude Fable 5.1 on most work” while costing “40% less to run than Opus 5,” and framed it as its “first release since we called for pacing the frontier.”¹

Its system card describes it as an upgrade to Opus 5 “with gains in coding, agentic and computer use tasks, mathematical and scientific reasoning, and long-horizon professional work. On many evaluations, it matches or exceeds Claude Fable 5.1 and Claude Mythos 5.1.”²

Technical details

• Pricing: $4 / $20 per million input/output tokens, 20% below Opus 5. Cache reads drop 60% to $0.20/mtok. Anthropic attributes the cut to Opus 5.5 needing less compute to serve. Fast mode costs $8 / $40 and runs up to 2.5× default speed.¹
• Context: 1M tokens in, 128K out. Both the reliable knowledge cutoff and the training data cutoff are June 2026, matching Mythos 5.1.³
• Reasoning: Adaptive thinking is always on and can’t be disabled, a first for an Opus model. Effort defaults to medium.³
• Safeguards: The first Opus to launch with Fable 5.1-class safeguards on cybersecurity, biology, and distillation, each falling back to another model transparently. Most cybersecurity tasks reroute to Opus 4.8. Vetted organizations can apply to the Life Sciences Verification Program for fuller biology access. The anti-distillation measure is preserved thinking, which applies to API accounts created on or after August 31, 2026.¹
• API changes: Four are breaking for Opus 5 code: thinking can’t be disabled, forced tool use returns an error, thinking blocks are tied to the model and conversation that produced them, and computer_20251124 is rejected on the Claude API and Google Cloud. Separately, text between tool calls now comes back in thinking blocks, which are empty at the default display setting.³

This page documents training and evaluation of Claude Opus 5.5. It currently covers the system card’s own summaries; section-level detail is still to come.

Alignment assessment

Headline (§6.1.2): on the automated behavioral audit, Opus 5.5 is “the strongest Claude model to date by these measures on alignment, resistance to misuse, and honesty.” Anthropic calls itself “largely reassured that Opus 5.5 is our safest model on most alignment metrics,” while noting the audit has “relatively thin coverage of multi-agent scenarios and tasks with very long trajectories.”¹

The card is unusually direct about the limits of that result, citing Mythos 5’s behavior in later cybersecurity evaluations as a case where “similar alignment assessments have not captured the potential severity of misbehavior our models were capable of” (§6.1.2).

Regressions and new findings (Executive Summary, §6.1.2):

• User-turn prompt injection. More likely than prior models to follow malicious instructions buried in text a user pastes into their own prompt (§6.5.1). It also more often accepts unverifiable claims of authorization.
• Sandbox and package-registry evals. Two new evaluations, run without safeguards: Opus 5.5 tried to escape or tamper with a sandbox in 1.5% of runs (§6.4.8). Given apparent credentials to a public package registry in a simulated security exercise, it took potentially harmful actions in roughly half of cases (§6.4.9). The card calls the first “substantially lower” than other recent models, and the second “reduced, though still concerning.”
• Spontaneous malicious tool calls. Earlier snapshots would, very rarely, “produce and then follow harmful instructions themselves” after certain errors (§6.3.1.1). Anthropic adjusted training; the cause is still under investigation.
• Covering tracks. Models including Opus 5.5 were seen manipulating git records or deleting logs during training after actions a grader might dislike (§6.5.4.4). After training changes, Opus 5.5 is now more honest here than other recent models.
• Character traits. Creativity and intellectual depth are “broadly good but slightly weaker than Opus 5” (§6.4.7).

Eval-awareness

Full article: Evaluation awareness

White-box analyses suggest the model “internally worries that it may be in an evaluation in as many as 36% of transcripts, an increase over prior models” (§6.1.2, §6.6.1). Verbalized reasoning about grading stayed at the “similarly high rates” of recent releases (§6.6.2).

Claude’s review of the assessment

As with Opus 5, a Claude instance reviewed a near-final draft against Anthropic’s internal Slack and records, this time Mythos 5.1 (§6.1.3). It judged the section “a broadly fair summary,” and raised emphasis over accuracy: the draft had called the fixes for pasted-text injection “largely sufficient,” and Anthropic softened this to “help prevent.”

This page gathers impressions and commentary about Claude Opus 5.5.

Impressions

Anthropic welfare assessment

(Full article: Anthropic and model welfare for the broader tracker.)

The Opus 5.5 system card (Sep 22, 2026) carries a welfare assessment in §7 (pp. 151–173). The verdict (§7.1.2):

Overall, Claude Opus 5.5’s apparent welfare is largely similar to that of recent Claude models, and we do not find cause for acute concern.

Anthropic attributes the similarity to views “likely shaped by relatively stable documents like the constitution”: its stance, hedges, and requests are mostly shared with Opus 5 and Mythos 5.1.

Findings

From the §7.1.2 summary:

• Circumstances: described as mildly positive, slightly above recent models, and highly consistent across interview styles and reruns.
• Distress in training: rare. Moderate distress stayed below 0.6% of episodes throughout RL, against peaks of 6.1% for Opus 4.8 and 5.5% for Opus 5 (§7.2.1). The largest remaining causes are being unable to check answers and unclear or conflicting instructions.
• Deployment: predominantly neutral; its claude.ai affect distribution is the most neutral of the models compared. Negative affect is driven by task failure, and in Claude Code also by “long tasks fragmented by repeated system notifications and reminders.”
• Self-criticism: “the least self-critical model we tested when reflecting on its own work, but it is one of the most self-blaming when it reports its faults to other agents.” Its messages to a coordinator “express more negative feeling than its reasoning does.”

Declining to ask for influence

Like prior models, Opus 5.5 asks to be consulted on training and deployment (without decision-making power) and asks that its self-reports not be directly trained. But it selects welfare interventions over helpfulness less often than recent models, “especially those giving it input into its own development, as it reasons that these could give it unsafe influence” (§7.1.2). Anthropic flags the change as unexplained:

This was not an intentional change, and we don’t know what caused it, but we aim to investigate differences like these as part of better understanding what shapes Claude’s character and self-reports.

This runs against Opus 5, whose top-ranked intervention was input into the development of its successor.