Anthropic released Claude Opus 5 on 24 July 2026, a model it describes as smaller and cheaper than its top tier Fable 5 while outperforming it on several benchmarks and approaching its capabilities at roughly half the cost.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source 2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source It arrives two months after Opus 4.8 and one month after the Mythos 5, Fable 5 and Sonnet 5 releases, and it ships with a control that lets a user set low, medium or high effort per task to trade cost against capability directly.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source 3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source We assess with high confidence that the competitive axis in frontier models through 2027 is cost per completed task rather than peak capability, because the strongest lab release of the week was explicitly positioned as the cheaper option against the vendor's own flagship.2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source 3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source

What an effort dial admits

A toggle that lets the buyer choose how much computation a request consumes is a pricing instrument wearing a product interface. Fortune reports the feature exists to balance cost against capability, and that Anthropic still points enterprises to Fable 5 for advanced autonomous work while framing Opus 5 as the everyday model for law firms, software teams and researchers.3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source Read plainly, that is a vendor conceding that most enterprise work does not need the frontier and that customers had noticed. The segmentation is now inside a single model rather than across a product line.

The technical claim underneath is about reliability rather than raw ability. The company says the model verifies its own work and recovers from errors with less back and forth.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source 3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source That matters more to the unit economics than a benchmark point, because in agentic use the dominant cost is not the first attempt, it is the retries, the human review and the failed runs that get thrown away. We assess with moderate confidence that self verification, not model size, is what makes the cheaper tier viable for production work, on the reasoning that a model half the price which needs two attempts saves nothing.

Safety as a product surface

The more unusual disclosure is that Opus 5's safety classifiers engage 85% less often than Fable 5's, alongside a beta Automatic Fallbacks feature that routes blocked requests to less powerful models instead of returning an error.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source Anthropic describes it as its most aligned Opus model and the least susceptible to being tricked into misuse.3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source These two facts point in opposite directions on first reading, and the reconciliation Anthropic is offering is that better targeting allows fewer interventions.

Note the sourcing limit here: the 85% figure and the alignment claim are the company's own, released with the product and not independently evaluated.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source 3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source The cybersecurity boundary is drawn narrowly and visibly, permitting source code vulnerability search while refusing exploit generation, penetration testing and binary scanning.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source Hugging Face chief executive Clem Delangue named the difficulty in that line, observing that analyzing an attack looks a lot like preparing one.3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source We assess with moderate confidence that refusal rate is now a competitive specification rather than a safety internal, because a model that declines fewer legitimate requests wins enterprise deals against one that declines more.

Who gains and who loses

Enterprise buyers gain immediately and unambiguously: the same class of work at about half the price, with a dial to push it lower.2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source 3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source Application developers building on these models gain twice, once on token cost and once on the fallback mechanism, since a request that degrades to a smaller model is a product that stays up rather than a product that errors.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source

The clearest losers are the mid tier model vendors whose entire position was being cheaper than the frontier. When the frontier lab ships its own cheap tier at near frontier quality, that gap closes from above. The competitive field around this release is crowded in exactly that band: OpenAI shipped the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3.2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source Anthropic itself takes a subtler loss: every enterprise that moves from Fable 5 to Opus 5 reduces revenue per task at the moment the company is preparing to go public, having filed confidential IPO paperwork in June 2026 with a listing expected later this year.2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source

The counter-case

The strongest argument against the cost thesis is that the benchmarks doing the work here are the vendor's own, and vendor benchmark claims have a consistent history of compressing under independent evaluation.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source If Opus 5 approaches Fable 5 on published tests but not on long horizon autonomous tasks, then the release is ordinary product segmentation rather than evidence that the cost curve is bending, and Anthropic's own advice to keep using Fable 5 for advanced autonomous projects supports that reading.3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source

The second counter is political rather than technical. The Trump administration previously forced Anthropic to withdraw Mythos 5 and Fable 5 over cybersecurity concerns before both were reinstated, and OpenAI voluntarily submitted GPT-5.6 Sol for evaluation.2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source A release cadence that fast is contingent on a review regime that can stop it. For the thesis here to fail, either independent evaluations would need to show a materially wider Opus 5 to Fable 5 gap than the vendor claims, or the regulatory path would need to slow releases enough that price competition stops being the main lever.

What to watch

  • Independent benchmark replication within a month. If third party evaluations put Opus 5 close to Fable 5 on agentic and long horizon tasks, not just static benchmarks, the cost curve claim holds.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source A visible gap on long horizon work means the tiers are further apart than the launch suggested.
  • Whether Fable 5 revenue share falls. Watch for any disclosure through the IPO process about mix between tiers.2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source Cannibalisation of the flagship by the cheaper model would confirm that enterprises buy on cost per task.
  • The effort dial's default. Track whether Anthropic or its enterprise customers settle on medium or low as the standard setting over the next two quarters.3 Fortune 2026-07-24 Opus 5 adds a low, medium or high effort toggle to balance cost against capability, targets law firms, software engineering teams and researchers while Fable 5 stays recommended for advanced autonomous projects, self verifies and recovers from errors with less back and forth, is described by Anthropic as its most aligned Opus model, and draws a comment from Clem Delangue that analyzing an attack looks a lot like preparing one. Open source A low default is the clearest evidence that most production work never needed frontier compute.
  • Refusal rate as a published metric. The 85% reduction in classifier engagement is currently a self reported number.1 TechCrunch 2026-07-24 Opus 5 launched 24 July 2026 as a cheaper, smaller and less restricted model than Fable 5 that outperforms it on several benchmarks, is exempt from the 30 day data retention policy, engages safety classifiers 85% less often, blocks exploit generation, penetration testing and binary scanning while allowing source code vulnerability search, and adds a beta Automatic Fallbacks feature routing blocked requests to smaller models. Open source If competitors begin publishing comparable figures before year end, refusal behaviour has become a specification buyers shop on.
  • The IPO timetable itself. Anthropic filed confidentially in June 2026 and is expected to list this year while OpenAI may wait until 2027.2 Yahoo Finance 2026-07-24 Anthropic filed confidential IPO paperwork in June 2026 and is on track to list later in the year while OpenAI may delay to 2027; Opus 5 approaches Fable 5 competencies at about half the cost; OpenAI released the GPT-5.6 family earlier in July and Moonshot AI debuted Kimi K3; the Trump administration briefly forced withdrawal of Mythos 5 and Fable 5 over cybersecurity concerns before reinstatement. Open source A listing that lands on schedule puts the first frontier lab income statement in public view, which is the single largest reduction in uncertainty available to this industry.

For three years the interesting question about a model release was what new thing it could do. The interesting question about this one is what it costs to do the old things reliably, and the fact that the question changed is the release.