Anthropic Delivers Flagship-Tier AI Performance at a Discounted Rate Prioritizing Safety

On September 22, 2026, Anthropic and OpenAI both dropped new AI models with lower prices, but the two companies bet on opposite ends of their product lineups.

Anthropie released Claude Opus 5.5, a model designed to match the performance of its higher-tier Fable 5.1 while costing 40 percent less than the outgoing Opus 5. OpenAI, on the same Tuesday, cut API prices in half for GPT-6 Sol and Luna, the two models sitting below its flagship GPT-6 Astra. Astra itself didn’t get cheaper. The timing collision meant each company’s benchmark tables were assembled against the other’s previous generation, creating a strange scenario where the newest releases are being evaluated against outdated competitors.

Anthropic’s approach centers on pushing near-flagship capability down the pricing ladder. Opus 5.5 now costs $4 per million input tokens and $20 per million output tokens, a 20 percent drop from Opus 5’s $5 and $25. The deeper savings show up in cached reads, which Anthropic says dominate agentic and coding workflows. Those tokens fell from $0.50 to $0.20 per million, a 60 percent cut. When you combine the token price reductions with the fact that Opus 5.5 uses fewer tokens per task, the company says typical workloads end up costing about 40 percent less. There’s also a faster mode priced separately at $8 and $40 per million tokens for up to 2.5 times the speed, giving developers a way to trade cost for latency when they need it.

OpenAI went another direction. GPT-6 Sol dropped from $4 and $20 per million tokens to $2 and $10, exactly half. GPT-6 Luna dropped from $0.20 and $1.20 to $0.10 and $0.50, also half. Both cuts are measured against GPT-5.6’s promotional pricing rather than list pricing, which OpenAI states directly in its announcement. Astra, which launched on September 3 and remains OpenAI’s strongest model across the board, didn’t appear in the pricing table for this release at all.

So Anthropic priced down its second-highest tier and positioned the result as a way to buy Fable-class output at Opus prices. OpenAI priced down its two cheapest tiers and left the top of its lineup where it was, framing Sol and Luna as ways to bring Astra’s training techniques to everyday budgets rather than as replacements for the flagship.

Anthropic and OpenAI launched cheaper AI models on September 22, 2026, but their strategies differ. Claude Opus 5.5 brings near-flagship performance to a lower tier with major cached-token savings and detailed safety claims, while GPT-6 Sol and Luna offer deeper direct price cuts below GPT-6 Astra. Developers must weigh effective workload cost, access restrictions, safety transparency, and benchmark methodology rather than relying on headline prices or isolated scores.

Both companies cite results from AutomationBench, a Zapier-built test of business workflows across 47 connected apps. Two of the numbers match exactly. Anthropic’s table lists Claude Opus 5 at 26.9 percent and Claude Fable 5.1 at 31.4 percent. OpenAI’s table, published the same day, lists the identical figures for the identical models: Opus 5 at 26.9 percent and Fable 5.1 at 31.4 percent. For an industry where every company runs its own version of a benchmark under conditions the other side can’t fully see, two competitors independently reporting the same scores for the same models is a small but real point of confirmation.

Not every figure holds up as cleanly. Anthropic’s table puts GPT-6 Astra at 41.4 percent on AutomationBench. OpenAI’s own table, in the same release, puts GPT-6 Astra at 30.3 percent, run at “low” reasoning effort so its cost per task would be comparable to Sol. The gap isn’t necessarily a contradiction. Neither company states the effort setting Anthropic used when it recorded 41.4 percent, and AutomationBench scores on this kind of agent shift with reasoning effort more than most benchmarks. But it means the same model, on the same test, appears twice in the public record eleven points apart, and a reader comparing the two tables at face value would draw the wrong conclusion about where Astra stands.

A similar gap shows up on OSWorld 2.0, a computer-use benchmark. Anthropic’s table has Claude Opus 5 scoring 74.0 percent, partial credit, with no effort level specified for the competitor row. OpenAI’s release states that Claude Opus 5 at medium effort scores 60.3 percent on the offline version of the same benchmark, which it uses to show GPT-6 Sol at xhigh effort reaching a comparable 60.5 percent at a much lower cost. Fourteen points separate the two reported scores for the same Anthropic model on what both companies call the same test. Effort level again looks like the likely explanation, but neither release makes that explicit for the competitor’s number, which is exactly the kind of gap a reader should notice before citing either figure as settled.

Where each company’s own model is concerned, the numbers are more legible because effort settings are disclosed. On DeepSWE v1.1, a software-engineering benchmark, GPT-6 Sol at max effort scores 68.8 percent, which OpenAI places within 1.1 percentage points of Claude Fable 5’s best recorded score of 69.9 percent at xhigh effort, achieved at roughly 80 percent lower cost per task. GPT-6 Luna at max effort scores 66.6 percent, which OpenAI calls comparable to Opus 5 and Fable 5 at medium effort, at 93 and 96 percent lower cost respectively. A footnote on that comparison discloses that Fable 5’s score was used because Fable 5.1’s wasn’t available, a small but useful piece of transparency about what’s being measured against what.

Anthropic frames Opus 5.5 as its first release since chief executive Dario Amodei’s public call to pace frontier AI development against safety practices. The company reports that Opus 5.5 scored better than any recent Claude model on its automated behavioral audit, a suite of roughly 2,000 simulated scenarios, and that it attempted to cross containment boundaries around 85 percent less often than Opus 5 or Claude Mythos 5.1 in a new evaluation built for that purpose. Opus 5.5 launches with cybersecurity and biology safeguards similar to those on Claude Fable 5.1, meaning many cyber tasks route to an older model, Opus 4.8, unless the user is enrolled in a verification program. It also carries preserved thinking, an anti-distillation measure that blocks attempts to extract a model’s reasoning by editing its prior context.

OpenAI’s release for Sol and Luna states that both models show improvements over their GPT-5.6 predecessors on internal alignment evaluations, including lower rates of what it calls misleading claims about coding work, and points to a system card for the full results. The webpage itself displays chart placeholders for five evaluation categories: coding deception, broken search, reviewer bypass, warning circumvention, and unauthorized interaction, without rendering the underlying numbers in the page’s text. That leaves Anthropic’s safety claims for Opus 5.5 easier to check against a specific number than OpenAI’s claims for Sol and Luna, at least from the announcement page alone.

Anthropic’s emphasis on safety testing and transparency aligns with the company’s broader positioning. The 85 percent reduction in containment boundary attempts is a concrete claim that developers can factor into risk assessments. The preserved thinking feature addresses a specific concern about model distillation, where a smaller model is trained to mimic a larger one’s reasoning patterns. By blocking attempts to extract reasoning through context editing, Anthropic is putting a technical barrier in place against a known attack vector.

OpenAI’s safety disclosures are less granular in the public-facing materials. The chart placeholders suggest the data exists, but developers looking for specifics before integrating Sol or Luna into production systems will need to dig into the system card rather than pulling numbers directly from the announcement. That’s a practical difference for teams that need to compare safety profiles quickly.

The most important development is not that AI prices fell again; it is that pricing, capability, and evaluation are becoming harder to compare cleanly. Anthropic offers a compelling package for agentic workloads that benefit from caching and for teams that value visible safety metrics, while OpenAI is more aggressive at the low end. In either case, a small representative workload test is more reliable than treating either company’s benchmark table as a neutral ranking.

Availability is where the two releases diverge most practically. Claude Opus 5.5 is live now across Amazon Web Services, Google Cloud, and Microsoft Azure, and developers can call it immediately as claude-opus-5-5 on the Claude Platform. GPT-6 Sol and Luna are live in ChatGPT Work and in Codex today for Plus, Pro, Business, Enterprise, and Edu subscribers, callable in the API as gpt-6-sol and gpt-6-luna. Free and Go users get access only to Luna, and only in the desktop app. Neither Sol nor Luna is available yet in ChatGPT’s standard Chat mode, and OpenAI says the rollout inside ChatGPT Work and Codex is staggered through the day rather than immediate for every account.

That means Anthropic’s Opus 5.5 is broadly available to any developer who wants to start using it right now, while OpenAI’s Sol and Luna have a more tiered rollout that depends on subscription level and product. If you’re a free-tier OpenAI user, you’re limited to Luna in the desktop app. If you’re building on the API, you can call Sol and Luna, but you can’t access them in the standard ChatGPT interface yet. Anthropic’s launch is simpler: if you have access to the Claude Platform or one of the major cloud providers, you can use Opus 5.5 today.

Anthropic has said Claude Sonnet 5.5 and Claude Haiku 5.5 will follow “in the coming weeks,” carrying many of the same efficiency and safety changes down through the rest of its lineup. OpenAI’s release covers Sol and Luna but says nothing new about GPT-6 Terra, the middle tier between them, or about whether Astra’s pricing will move. The two companies have now each cut prices on part of their lineup within hours of each other twice in three weeks, counting Astra’s September 3 launch, and the AutomationBench and OSWorld discrepancies above are worth revisiting once both sides publish fuller methodology notes rather than announcement-page summaries.

These releases invite several shortcuts that can produce misleading cost or capability comparisons.

Both releases land in a market where AI token prices are under pressure. The LLM Token Expenditure Index fell below $1 to 97 cents per million tokens in early September 2026, marking a new low and exerting financial pressure on leading AI companies including OpenAI and Anthropic. Price erosion stems from competition with lower-cost open-source Chinese models and dynamic pricing strategies introduced by frontier AI labs, squeezing revenue while fixed compute costs remain.

This price decline risks training consumers to expect cheaper AI access, potentially eroding the pricing power of established AI model providers during a critical IPO period. OpenAI and Anthropic are both navigating this tension between maintaining margin and capturing market share. OpenAI’s decision to cut prices on Sol and Luna while leaving Astra untouched suggests the company is betting that developers who need the best performance will still pay for it, while developers working with tighter budgets or less demanding tasks will move down to the cheaper tiers. Anthropic’s decision to compress the gap between Opus and Fable suggests a different bet: that offering near-flagship performance at a lower price will pull in customers who might otherwise have gone with a competitor’s cheaper model.

The fact that both companies launched on the same day is probably a coincidence, but the way they positioned their releases reveals how each one thinks about competition. Anthropic is emphasizing capability parity with its top tier at a lower price, backed by concrete safety improvements. OpenAI is emphasizing cost reduction on models that bring Astra’s techniques to a broader audience, without undermining Astra’s premium position.

For developers choosing between the two, the decision comes down to what matters more: Anthropic’s transparent safety testing and broad immediate availability, or OpenAI’s deeper price cuts on the lower tiers with a more complex rollout. If you’re running high-volume agentic workflows where cached token reads matter, Anthropic’s 60 percent cut on those tokens could tip the balance. If you’re looking for the absolute lowest cost per token and don’t need flagship performance, OpenAI’s Luna at $0.10 and $0.50 per million tokens is hard to beat.

Both companies are also racing to fill out their lineups. Anthropic’s promise of Sonnet 5.5 and Haiku 5.5 in the coming weeks means the company is pushing the same efficiency and safety improvements across all three tiers. OpenAI’s silence on Terra and Astra pricing means those models are still positioned where they were before Sol and Luna launched, which could change if Anthropic’s next releases put more pressure on OpenAI’s pricing structure.

For teams already building on one platform or the other, these releases are mostly good news. You’re getting more performance per dollar, whether that’s through Anthropic’s efficiency gains or OpenAI’s direct price cuts. For teams evaluating which platform to choose, the decision gets more complicated. You’ll need to look at your specific workload, how much of it involves cached token reads, whether you need the absolute best performance or can trade some capability for lower cost, and how much transparency you need around safety testing.

The benchmark discrepancies are worth paying attention to. When two companies report different scores for the same model on the same test, it’s a signal to check the methodology before making a decision based on those numbers. Effort levels, test versions, and evaluation settings all matter, and right now those details are scattered across announcement pages, system cards, and footnotes rather than presented in a single comparable format.

The coming weeks will likely bring more clarity as both companies publish fuller methodology notes, other researchers run independent benchmarks, and developers start reporting real-world performance. Until then, the safest bet is to treat the announced numbers as directional rather than definitive, and to run your own tests on the workloads that matter to your specific use case.