Claude Opus 5 is framed as Anthropic’s everyday frontier model: cheaper than Fable 5, broadly accessible, and strong enough to be the default for general reasoning while still lagging Fable on best-in-class coding and some long-horizon tasks. Its standout claim is visual reasoning: one-shot prompts reportedly produced structurally plausible Brooklyn Bridge and game demos, with Arc AGI 3 scoring jumping to 30% from GPT 5.6 Soul Max’s 7.8/7%. The tradeoff is cost, token burn, and some long-duration muddiness; the recommendation is to use medium effort, not extra high, for better output. Safety is presented as looser than Fable 5 but still bounded: fewer false flags, more permission for benign cyber/bio work, no help on nuclear warhead requests, and stronger resistance to prompt injection.
Opus 5 is positioned as the model just below Claude Fable 5 after Opus 4.8, not the best, smartest, cheapest, or fastest, yet still a favorite of the year.
Opus 5 is described as costing as cheap as Opus 4.8 while being a massively different model, with visual intelligence that can break down pixels and reconstruct what it sees from code.
A one-shot Brooklyn Bridge demo reportedly took about 1.5 hours and around 150,000 tokens, with no post-prompt clicking, yet produced a visually rendered, structurally accurate bridge scene from a screenshot and instructions.
That Brooklyn Bridge prompt reportedly generated preset shot angles like a sunset wide shot and “a tower over here,” showed the bridge swaying, and rendered views under the deck while preserving structural soundness.
A one-shot custom-code game demo reportedly used no external assets, looked like Call of Duty, and was playable in a browser with textures, a skybox, footsteps, hit boxes, and weapon-firing audio features if audio were enabled.
A snowboard-game demo is described as visually indistinguishable from SSX Tricky and more impressive than SSX Tricky itself.
Signal’s X post is relayed as saying that if that snowboard game had shipped in the App Store in 2012 it could have grossed $5 million to $10 million over its lifespan, framing Opus 5 as wiping out roughly $10 million of value in 15 years.
A recommended AAA-visual workflow is to fan out sub-agents, split components among them, loop until a harsh critic sub-agent approves the work as AAA, and use a manager role to coordinate the agents.
The same sub-agent loop is claimed to generalize beyond games to Excel work and architecture blueprints.
Arc AGI 3 is called the hardest AI-model test, and Opus 5 is said to score 30% versus GPT 5.6 Soul Max’s prior high of 7.8, with Arc describing Opus 5 as converting layouts into algebraic notation before reasoning.
Opus 5 is said to score four to five times the next leading model on visual and logical reasoning while costing a fraction of GPT 5.6 Soul Max.
Compared with Fable 5, Opus 5 is claimed to do most of the same things at 50% less cost, while Fable 5 pricing is given as about $5 per million input tokens and $25 per million output tokens, or $30 overall in usage credits.
Opus 5 is described as the default versus Opus 4.8 and as triggering fewer safety flags than Fable 5, including on benign questions like whether mitochondria are the powerhouse of the cell.
On bio and cybersecurity prompts, Opus 5 is claimed to “just do the thing,” resist prompt injection because it is extremely aligned, and still refuse direct nuclear-warhead instructions while allowing benign OS bug-checking.
Anthropic is said to have shipped four frontier models in eight weeks—Mythos 5, Fable 5, Sonnet 5, and Opus 5—while Fable 5 is characterized as a long-thinking planner that spawns sub-agents, checks plans, and tends to improve with more runtime.
Opus 5 is framed as the workhorse for shorter tasks, but long thinking can make it messier; the recommendation is to use medium effort instead of extra high because it thinks less long.
Plan access is said to include Opus 5 on the free tier, substantial access on the $20/month plan, and Fable 5 on the $100 max plan, with $100 in usage tokens being handed out for testing.
Opus 5 is recommended as the everyday model for general reasoning and most general questions, while Fable remains better for best-in-class coding, some health use cases, and certain cybersecurity and bio evals.
The launch is described as the first Anthropic/frontier-lab model release that did not advertise how intelligent it is, possibly out of fear of regulation.
AGI Index is said to have had AGI 1 and AGI 2 hit 100% as models improved, while AGI 3 is the newer, harder benchmark where GPT 5.6 scored 7% and Opus 5 scored 30%.
Opus 5 is claimed to natively spin out multiple AI agents in parallel, making it a preview of how work will be executed.
The sponsor pitch is Ledger Agent Stack: four open-source tools for Ledger Wallet, Ledger Enterprise, and Ledger Enterprise multisig that let AI agents draft transactions but require Ledger signer approval before execution, with support for cloud code, Codex, and Cursor.
A regulatory and safety story says the US government got involved with Fable 5, called it too powerful, required more safeguards, and forced Anthropic to pause release without safety guards before iterating.
Opus 5 is presented as an embellishment of a good version of Mythos 5, with safeguards around the same as standard models, fewer blocks than Fable, and prompt-injection protection that still allows benign cybersecurity work but not direct exploit-defense creation.
A genomics example says Fable 5 could not solve it, while Opus 5 handled a breakthrough involving genomic sequencing and skin aging reversal.
Opus 5 is described as accessible to everyone, a major upgrade, a token guzzler, 50% cheaper than Fable, and more verbose than expected because it inherited Fable’s “speaks a lot” style.
Kimmy K3 is cited as a recent Moonshot Labs model whose weights should become open, frontier-capable at a fraction of cost but lacking safeguards, alongside mention that Nvidia and Microsoft signed an open-source waiver this weekend while Anthropic did not.