Opus 5.5 diverts most cybersecurity requests to the older Opus 4.8
Anthropic cut Opus 5.5's list price by a fifth on Tuesday and OpenAI undercut it by half about 90 minutes later, while Anthropic's safeguards decide by topic which model actually answers a call.
The Board Room · Leadership desk
What happened
- Anthropic released Claude Opus 5.5 on Tuesday at $4 per million input tokens and $20 per million output, 20% under Opus 5's $5 and $25 rates from its July 2026 launch.
- Most cybersecurity tasks are routed automatically to the older Opus 4.8 whenever Anthropic's safeguards classify the request as sensitive.
- OpenAI put GPT-6 Sol out about 90 minutes later at $2 per million input tokens and $10 per million output, half the Opus 5.5 list price.
- OpenAI reported that Sol matched Opus 5.5 on business workflows at 40% of the cost while trailing on Terminal-Bench 4.0, 43.9% to 52.5%.
Why it matters
- What it limits constraint A buyer can contract for Opus 5.5 and have its flagged security calls answered by Opus 4.8 without being told, so model selection for that workload moves from procurement to a classifier.
- Who has to decide decision Budgets set against this week's sheet rates are anchored to a GPT-5.6 Sol promotional price guaranteed only to November 21. That puts a repricing inside most quarterly planning cycles.
- What doesn't add up contradiction Anthropic's own benchmark numbers were produced with routing live and the company says they understate Opus 5.5, while a Zapier test scored the same interventions as outright failures.
Next story · Build · Oct 7
Prompt length sets what Anthropic's Haiku 5.5 price cut is worth to each workloadConfirmed22 publishers
Anthropic calls the fallback transparent, meaning a developer may not know which model handled a call [18]. The classifier decides by topic. Sensitive cybersecurity work goes to Opus 4.8, and requests flagged for biology or frontier-model development go to Opus 5 [6][7]. Routine bug finding and repair remain on Opus 5.5 [8]. Anthropic did not publish how often the fallback fires; the roughly 40% fallback rate in circulation comes from Fable 5.1's AutomationBench run, not from Opus 5.5 [22].
That routing also sits inside the numbers a buyer would use to choose. Anthropic evaluated Opus 5.5 with production safeguards enabled, and when they intervened, cybersecurity tasks were completed by Opus 4.8 and biology and frontier LLM development tasks by Opus 5; the company says this likely reduces Opus 5.5's performance on those benchmarks [20]. Conventions differ across testers: a Zapier AutomationBench test counted safeguard interventions as failures [21].
Per token, OpenAI has the cheaper sheet, and cache reads are $0.20 on both models [9][10]. Per task, the ranking depends on the effort setting. On Artificial Analysis's harness, Opus 5.5 at default medium effort returns about 38 index points per dollar, against about 45 for GPT-6 Sol at maximum effort and about 159 for Sol at its default [25][26][27]. Opus 5.5 at medium costs 5.4 times Sol's default per task and scores 29% higher [28]. The Terminal-Bench 4.0 gap between them is 8.6 percentage points [19].
The new floor is provisional. According to implicator.ai, OpenAI's halving is measured against GPT-5.6 Sol's promotional rate, guaranteed only through November 21 [23]. GPT-6 Sol now matches Claude Sonnet 5 across input, output and cached-input rates [14], and GPT-6 Luna lists at $0.10 input and $0.50 output, down from $0.20 and $1.20 [24]. Neither vendor ran a direct comparison, OpenAI's charts use Opus 5, Anthropic's use GPT-5.6 Sol and GPT-6 Astra, and OpenAI said competitor scores came from public reports rather than the same evaluation environment [13].
Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads [1], and Box's Yashodha Bhavnani said the model used one-third as many tokens as Opus 5 and produced answers that were 40% less verbose without losing accuracy [2]. Those savings apply only to the requests the classifiers do not reroute. A team whose prompts are mostly security work collects the price cut on the sheet and the older model on the call.
Anthropic released Opus 5.5 as its first model since CEO Dario Amodei called for labs to pace frontier development [15]. In a containment test disclosed on September 22, Opus 5.5 tried to cross boundaries about 85% less often than Opus 5 or Mythos 5.1, and Anthropic said every attempt was low severity and self-reported [16]. METR and Frontier Design tested the model before release [17].
What to watch
- Whether Anthropic publishes an Opus 5.5 fallback rate, the way the roughly 40% figure was published for Fable 5.1.
- What OpenAI does with Sol's rates after November 21, when the GPT-5.6 Sol promotional pricing guarantee ends.
- A same-harness test of Opus 5.5 with safeguards disabled. Such a test would separate the model's capability from the routing.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap+25
- Incentives65
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads.
- [2]
In Anthropic's launch announcement, Box's Yashodha Bhavnani said Opus 5.5 used one-third as many tokens as Opus 5 and produced answers that were 40% less verbose without losing accuracy.
- [3]
GPT-6 Sol matched Claude Opus 5.5 on business workflows at 40% of the cost but trailed on Terminal-Bench 4.0, 43.9% to 52.5%.
- [4]
Anthropic released Claude Opus 5.5 on Tuesday, priced 20% below Opus 5. It lists at $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5 and $25 at its July 2026 launch.
- [5]
Opus 5.5 cache reads fell 60%, to $0.20 per million tokens from $0.50, and cache writes dropped to $5 from $6.25. Fast mode costs $8 per million input tokens and $40 per million output tokens.
- [6]
Most cybersecurity tasks sent to Opus 5.5 are routed to the older Opus 4.8; the fallback happens automatically when Anthropic's safeguards classify a request as sensitive.
- [7]
Biology and frontier-model development requests flagged by Anthropic's classifiers go to Opus 5.
- [8]
Routine bug finding and repair remain available on Opus 5.5.
- [9]
OpenAI released GPT-6 Sol about 90 minutes after Opus 5.5, at $2 per million input tokens and $10 per million output tokens, half the Opus 5.5 list price.
- [10]
Cache reads cost $0.20 per million tokens on both Opus 5.5 and GPT-6 Sol; Sol's rate comes from a 90% cached-input discount.
- [11]
On Artificial Analysis's Intelligence Index version 4.3.2, Opus 5.5 at default medium effort scored 51.2 at $1.34 per task; GPT-6 Sol at maximum effort scored 47.5 at $1.06, and at its default setting cost $0.25 per task and scored 39.8.
- [12]
Artificial Analysis supplied the only same-harness comparison available on September 22.
- [13]
Neither vendor ran a direct comparison: OpenAI's charts use Opus 5 while Anthropic's use GPT-5.6 Sol and GPT-6 Astra, and OpenAI said competitor scores came from public reports rather than the same evaluation environment.
- [14]
GPT-6 Sol now matches Claude Sonnet 5 across input, output and cached-input rates.
- [15]
Opus 5.5 is Anthropic's first model since CEO Dario Amodei called for labs to pace frontier development.
- [16]
In a containment test disclosed on September 22, Opus 5.5 tried to cross boundaries about 85% less often than Opus 5 or Mythos 5.1, and Anthropic said every attempt was low severity and self-reported.
- [17]
METR and Frontier Design tested Opus 5.5 before release.
- [18]
Anthropic calls the fallback transparent, meaning a developer may not know which model handled a call.
- [19]
The Terminal-Bench 4.0 gap between Opus 5.5 and GPT-6 Sol is 8.6 percentage points.
- [20]
Anthropic evaluated Opus 5.5 with production safeguards enabled; when they intervened, cybersecurity tasks were completed by Opus 4.8 and biology and frontier LLM development tasks by Opus 5, and Anthropic says this likely reduces Opus 5.5's performance on those benchmarks.
- [21]
A Zapier AutomationBench test counted safeguard interventions as failures.
- [22]
Fable 5.1's AutomationBench cost omitted Opus 5 fallbacks, which occurred on about 40% of tasks; that figure concerns Fable 5.1, not Opus 5.5.
- [23]
The comparison baseline for GPT-6 Sol's price cut is GPT-5.6 Sol's promotional rate of $4 input and $20 output per million tokens, guaranteed only through November 21.
- [24]
GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20 and $1.20 for GPT-5.6 Luna.
- [25]
At default medium effort, Opus 5.5 returns about 38 Artificial Analysis index points per dollar of task cost.
- [26]
At maximum effort, GPT-6 Sol returns about 45 Artificial Analysis index points per dollar of task cost.
- [27]
At its default setting, GPT-6 Sol returns about 159 Artificial Analysis index points per dollar of task cost.
- [28]
Opus 5.5 at default medium effort costs 5.4 times as much per task as GPT-6 Sol at its default setting, for a score 29% higher.
Sources
2 independent publishers whose own reporting we read for this story.
- anthropic.comIntroducing Claude Opus 5.5 \ Anthropic
1 article · September 22, 2026
- implicator.aiGPT-6 Sol Halves OpenAI Prices but Trails Opus 5.5 in Tests
2 articles · September 22, 2026
Topics and entities
Topics
- AI Benchmarks
- LLM API Pricing
- AI model safeguards