Three launches in three weeks: what shipped

Between September 22 and October 13, 2026, Anthropic, OpenAI and Google each shipped a new frontier model, and Anthropic refreshed the rest of its lineup inside the same window. If you run AI in production, that's a month of evaluation work arriving at once, and a pricing sheet that changed while you were reading it.

Here is what shipped, according to the vendors.

Claude Opus 5.5 arrived on September 22. Anthropic prices it at $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, and says it costs about 40% less on typical workloads. Anthropic reports 66.4% on Terminal-Bench 4.0, against 52.3% for Opus 5. It's available on AWS, Google Cloud and Azure.

Claude Sonnet 5.5 followed on September 28 at $2 / $10, with Anthropic claiming it costs "up to 30% less per task" than Sonnet 5. Claude Haiku 5.5 came on October 7, starting at $0.10 / $0.50 with a 1M-token context window. Anthropic's model overview positions it for classification, routing, extraction and subagent work.

GPT-6.1 Sol shipped on September 29. OpenAI's developer changelog lists it at $2 input, $0.10 cached input and $10 output per million tokens, with multi-agent support (delegating work to subagents) in beta in the Responses API. An "Ultrafast" mode followed on October 8. The system card describes capabilities close to GPT-6 Astra with a focus on speed and cost, and OpenAI rates the model "Critical" for cybersecurity under its Preparedness Framework.

Gemini 4 Argon became generally available on October 13. Google's launch post gives introductory pricing of $2 / $10 until November 30, 2026, then $4 / $20, and a "1M unified input-output token limit". It's available through AI Studio, the Gemini platform and Antigravity.

Price versus capability: the 2026 tiers

Put the list prices side by side and a pattern shows up.

Model Released Input / output per 1M tokens
Claude Haiku 5.5 Oct 7 from $0.10 / $0.50
Claude Sonnet 5.5 Sep 28 $2 / $10
GPT-6.1 Sol Sep 29 $2 / $10 ($0.10 cached input)
Gemini 4 Argon Oct 13 $2 / $10 until Nov 30, then $4 / $20
Claude Opus 5.5 Sep 22 $4 / $20

Three of these models sit at exactly the same list price, and one of them only temporarily. If you benchmark Argon in November and choose it on cost, redo the sums at $4 / $20, because that's the price from December. The output rate doubles overnight, and for output-heavy workloads that's the number that decides the bill.

The table also shows how wide the spread has become. Haiku 5.5's input price is a fortieth of Opus 5.5's. That gap is the real design opportunity this quarter. Most steps in an agent pipeline aren't frontier-hard, and the ones that are can be identified with data.

List price is only half the comparison, though. Per-task cost depends on how many tokens a model needs to finish the job, how often it needs a retry, and how much of your prompt is cacheable. Sol's cached input at $0.10 is a twentieth of its normal input rate, and the Claude docs say prompt-cache reads are billed at a fraction of the base input price (10% on Haiku 5.5). If your workload resends a long system prompt or document context on every call, caching moves the comparison more than the headline rate does. Anthropic's "about 40% less on typical workloads" for Opus 5.5 is also a per-task claim, and per-task claims are exactly the ones you need to check on your own tasks.

Benchmarks are self-reported, so run your own evaluations

Every figure above comes from the company selling the model. That doesn't make the figures wrong. It makes them a hypothesis.

Terminal-Bench 4.0 measures something real (an agent working through tasks in a terminal), and a 14-point jump is large. But your workload isn't Terminal-Bench. A tax research question, a product-page audit, a support-ticket triage: each fails in its own ways, and the only benchmark that predicts how a model will handle them is a set of your own cases with known-good answers.

If you don't have one yet, this is what we'd build. Take 100 to 300 real tasks from production, anonymized, each with a grader. Some graders are exact (did it extract the right invoice total?). Some are rubric-based and scored by a model, with a human reviewing a sample on every run so you notice when the model grader drifts. Track accuracy, cost per task, latency at p50 and p95, and refusal rate. When a new model ships, you run the harness the same afternoon and have a decision by Friday.

Designing for model churn: routing, abstraction, fallbacks

Three frontier launches in three weeks is roughly the pace now. Your architecture should assume the best model for any given step will change several times a year.

In practice that comes down to a few decisions. The first is a single internal interface for model calls, so changing provider is a config change and not a refactor. We built Promptify's backend around that idea: one execution engine in front of several LLM providers, so a prompt template can move from one model to another without touching the product.

The second is routing per step rather than per application. TaxFigure runs a multi-model stack (its tech list includes GPT-4o, Gemini Flash, Perplexity Sonar, DeepSeek and Cohere Rerank), and a research coordinator splits each question into parallel sub-tasks for specialized agents. Different jobs in the pipeline, different models.

The third is fallbacks that have been tested, not merely configured. When your primary provider has an outage or starts rate-limiting you, the fallback should be a model your harness has already scored on that exact step.

Prompts are the quiet lock-in. A prompt tuned for one model family often does worse on another. Version prompts per model, and let the harness tell you when a prompt needs rework for a new model, instead of finding out from a customer.

Build vs. buy: when a single-vendor stack makes sense

Provider-agnostic design has real costs. You maintain an abstraction layer, your eval matrix grows, and you sometimes give up a vendor-specific feature such as Sol's multi-agent beta in the Responses API.

A single-vendor stack is a reasonable choice when you're early and speed matters more than optionality, when one vendor's particular feature is central to your product, or when procurement decides for you. That last argument is weaker than it used to be. Opus 5.5 is available on all three major clouds, so you can often buy the model you want through the cloud contract you already have.

What we'd avoid is accidental lock-in: SDK calls scattered across the codebase, prompts without version history, no eval set. That's the stack that can't take advantage of a 40x price spread when one appears.

What we'd change in a production stack this quarter

For a team already running agents in production, this is the short list:

  • Run your eval harness against Opus 5.5, Sonnet 5.5, Sol and Argon on every step that currently uses a frontier model.
  • Test Haiku 5.5 on each classification, routing and extraction step. Those are the jobs Anthropic designed it for, and they're where the price gap is widest.
  • Model Argon's cost at its post-November price, not its launch price.
  • Read the GPT-6.1 Sol system card before giving it tools with security-relevant access. A "Critical" cybersecurity rating belongs in your threat model.
  • Check cache hit rates per prompt template. With cached input priced at a small fraction of normal input on both Claude and Sol, restructuring a prompt can save more than switching models.

None of these takes more than a few days if the harness and the routing layer exist. If they don't, that's the first thing to build, and it's the kind of work the integration and testing stage of our AI deployment program exists for.

Sources