60% Cheaper: Anthropic's Opus 5.5 Takes On OpenAI's New GPT-6 Models

Anthropic cut prices on its flagship coding and knowledge-work model by as much as 60 percent this week, even as it claimed modest performance gains over rivals, according to Ars Technica. The release, called Opus 5.5, landed alongside OpenAI's own cost-focused update, GPT-6 Sol and Luna, in a pairing Ars Technica frames as evidence the "frontier AI model race has entered its comparison shopping phase."

Neither company is pitching the releases as capability leaps. Instead, both are betting that enterprise customers — the buyers who actually pay the bills for coding assistants and agentic workflows — care more about the price per token than about incremental gains on a benchmark leaderboard.

What did Anthropic and OpenAI announce this week?

Anthropic's Opus 5.5 is the newest version of the model the company positions as its mass-market workhorse for coding and complex knowledge work, according to Ars Technica. OpenAI, meanwhile, rolled out GPT-6 Sol and Luna, the latest entries in the GPT-6 family aimed at the middle and lower end of its lineup rather than its top-tier model, GPT-6 Astra, which OpenAI released earlier in September.

Ars Technica reports that Anthropic is "playing a bit of catch-up" against OpenAI, since Astra has "sometimes been modestly beating Opus 5 in benchmarks and user sentiment." Anthropic's own benchmarks now show Opus 5.5 edging past Astra in some coding and knowledge-work tests, though the outlet characterizes the gains as modest.

How much cheaper is Opus 5.5 than its predecessor?

The pricing cuts are the headline. Anthropic's announcement, quoted by Ars Technica, lays out the numbers directly:

"Input and output tokens are $4 and $20 per million, 20% less than Opus 5. Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5. Opus 5.5 also generates output more than 30% faster than Opus 5."

Anthropic told Ars Technica the real-world savings run even deeper — closer to 40 percent for typical workloads at default settings — because Opus 5.5 also completes tasks using fewer tokens overall, compounding the per-token discount.

By the numbers

• Input tokens: $4 per million (down 20% from Opus 5)
• Output tokens: $20 per million (down 20% from Opus 5)
• Cache reads: $0.20 per million tokens (down 60% from Opus 5)
• Output speed: more than 30% faster than Opus 5
• Typical workload savings: roughly 40% versus Opus 5, per Anthropic

What do GPT-6 Sol and Luna add to OpenAI's lineup?

OpenAI's update is described by Ars Technica as "an iterative step forward" rather than a redesign. The company introduced its Sol, Terra, Luna naming scheme with the GPT-5.6 family and has since layered GPT-6 Astra on top as its most powerful and most expensive option. Sol and Luna sit below Astra, aimed at efficiency and speed rather than raw capability — the same positioning Anthropic is chasing with Opus 5.5's cache-read discounts and faster output.

Why are AI labs racing on price instead of raw power?

Ars Technica ties the pricing push directly to how enterprise buyers are behaving. Organizations have been experimenting with model routers — tools that automatically send simpler requests to cheaper models and reserve expensive frontier models for harder tasks — and weighing open-weight alternatives that undercut proprietary pricing entirely. Both Anthropic and OpenAI, the outlet notes, are "racing to compete with open-weight models" as that behavior spreads. A photo caption accompanying the Ars Technica piece captures the dynamic bluntly: "Developers and folks working on enterprise AI deployments are comparing these two. Cost is a big part of that."

What safety limits apply to Opus 5.5's capabilities?

Anthropic flagged that Opus 5.5 is "notably capable" in what it calls high-risk areas, including cybersecurity and biology, according to Ars Technica. The company said the same safeguards that applied to its earlier Fable 5.1 model carry over here: requests flagged as touching protected territory get automatically and transparently routed to an older, more restricted model rather than answered directly by Opus 5.5.

What should enterprise buyers watch next?

• Whether cache-read discounts and reduced token usage translate into real savings at scale, not just in vendor benchmarks
• How GPT-6 Astra's benchmark edge over Opus 5 holds up against the newly upgraded Opus 5.5
• Whether open-weight models and routing tools keep pulling workloads away from both companies' proprietary pricing tiers
• Any expansion of automatic routing safeguards as models grow more capable in flagged risk categories

For now, neither company is promising a leap in what these models can do. The pitch, as Ars Technica frames it, is doing largely the same work for markedly less money — a signal that the next phase of the AI model race may be fought on invoices rather than leaderboards.