The September 2026 model wave: what changed for builders and how to route by cost per task

What did each vendor actually release in September 2026?

Anthropic released one generally available flagship (Fable 5.1), a restricted sibling (Mythos 5.1) and two cheaper 5.5 models. OpenAI released GPT-6 Astra, then the cheaper Sol and Luna tiers, and previewed GPT-6.1 Sol. Google announced Gemini 4 Argon, which is not generally callable yet. The table summarises what each vendor published.

Sources: Anthropic’s launch pages, MacRumors and InfoQ for OpenAI, and Google’s Argon announcement. We could not confirm GPT-6 Astra’s API pricing or context window from a primary OpenAI page, so we have left those cells blank rather than repeat secondary-site figures.

The claims worth noting

• Fable 5.1. Anthropic says typical workloads cost around 25% less than on Fable 5, and highly agentic work up to around 45% less. It reports Terminal-Bench 4.0 at 55.8% (Fable 5: 42.0%) and Humanity’s Last Exam at 60.9% without tools and 65.0% with tools.
• Opus 5.5. Anthropic says it costs 40% less than Opus 5 on typical workloads, generates output more than 30% faster, and performs at the level of Fable 5.1 on most work.
• Sonnet 5.5. Anthropic claims 30%+ faster output and up to 30% lower cost for most work than Sonnet 5; it is also available on AWS, Google Cloud and Azure.
• GPT-6 Sol and 6.1 Sol. OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol in its internal testing, and that 6.1 Sol approaches Astra on several evaluations at one-fifth of Astra’s token prices (as reported by InfoQ).
• Gemini 4 Argon. Google raised the output limit to 1M tokens from 64K and reports 77.9% on DeepSWE v1.1, 51.3% on AutomationBench and 68% on CWE-bench v1, calling it its most resilient model yet against indirect prompt injection.

Why is list price the wrong number to compare?

Because agent workloads re-send the same context many times. A loop that re-reads a long system prompt, tool definitions and retrieved documents on every step pays mostly for input, and every vendor now discounts cached input heavily. Effort levels and speed tiers add two more dials. The real unit of cost is a completed task.

Caching is now the biggest lever

Anthropic cut Fable 5.1 cache reads by 75% to $0.25 per million tokens. Opus 5.5 cache reads cost $0.20, which Anthropic says is 60% less than Opus 5. OpenAI advertises 90% off cached input reads on GPT-6 Sol, and Google lists 95% off for Argon. But on Opus 5.5, cache writes cost $5 per million tokens against $4 for ordinary input, so caching only pays when the prefix is reused.

Effort and speed are explicit cost dials

Fable 5.1 and Mythos 5.1 expose five effort levels — Low, Medium, High, XHigh and Max — that trade reasoning depth for tokens and latency. Opus 5.5 has a Fast mode at double the standard price ($8 / $40). According to BGR’s DevDay coverage, OpenAI’s new Ultrafast tier runs up to 8x faster in Codex and up to 6x faster in the API at six times the standard API price, initially for GPT-6 Astra on Pro 500 and Enterprise plans. In code, all three dials belong in configuration, per task type.

What are gated models like Mythos 5.1 and Argon’s Fairwind access?

Both Anthropic and Google now release their strongest cyber-capable models to vetted organisations first. Claude Mythos 5.1 is limited to a set of US organisations through Anthropic’s Cyber Verification and Life Sciences Verification programmes. Gemini 4 Argon goes first to trusted cyber defenders through Google’s Fairwind Program, with wider API access promised but undated.

Google’s September roundup shows the same pattern one tier down: Gemini 3.8 Flash Cyber, pitched at autonomous vulnerability discovery and code patching, also goes through Fairwind, beside a generally available Gemini 3.8 Flash. Per Wikipedia, OpenAI’s initial GPT-6 rollout also used a restricted version that rejects certain cybersecurity prompts.

What this means in practice

• Do not design a product around a model you cannot call. As of 7 October 2026, Argon has no public API date; build on Gemini 3.8 Flash or another available model and treat Argon as an upgrade path.
• Defensive security teams should apply to the vendor programmes early; access, not price, is the bottleneck.
• Argon’s $2 / $10 price is introductory and rises to $4 / $20; Google has not said for how long, so model your costs at the higher rate.

Anthropic also announced Enterprise Frontier Safeguards: misuse-monitoring data is stored in the customer’s own cloud account rather than on Anthropic’s systems, so enterprises can keep a zero-data-retention agreement. It rolls out in phases from autumn 2026 at no separate charge beyond cloud costs.

How should you evaluate these models on your own tasks?

Build a small, versioned evaluation set from your real work and run every candidate model through the same harness. Vendor benchmarks are self-reported, use different harnesses and settings, and rarely match your prompts or tools. Score each model on success rate, cost per successful task and latency, then decide with your own numbers.

1. 01Collect 50–200 real tasks. Pull them from production logs, including the awkward ones, and record what a correct outcome looks like.
2. 02Fix the harness. Same system prompt, tools, retrieval and step limits for every model. Change one variable at a time.
3. 03Measure three numbers. Task success rate, total tokens and cost per successful task (including cache writes and retries), and p95 latency.
4. 04Sweep the dials. Test each model at two or three effort levels; a mid-tier model at high effort can beat a flagship at low effort on price and quality.
5. 05Re-run on every change. New model versions, prompt edits and tool changes all move results. Keep the eval in CI so regressions show up before users see them.

What does a practical model-routing strategy look like?

Route each task type to the cheapest model that meets its quality bar, and escalate only when needed. In late 2026 that usually means a mid-tier model by default, a flagship for the hardest steps, and a low-cost tier for classification and extraction — with caching, budgets and a fallback provider designed in from the start.

Implementation details that matter

• Order prompts for the cache. Stable content first (instructions, tool schemas, reference documents), variable content last. One changing timestamp at the top can wipe out the discount.
• Track cache-hit rate as a metric. Log cached versus uncached input tokens per request; if the ratio drops after a deploy, someone broke the prefix.
• Escalate on signals, not guesses. Promote a task to a stronger model or higher effort on failed validation, tool errors or failing tests.
• Plan fallbacks across providers. Keep a second provider configured per tier so an outage degrades service instead of stopping it, and evaluate the fallback too.
• Watch deprecation calendars. Pin model ids explicitly and subscribe to each vendor’s deprecation notices; releases this frequent mean older ids retire quickly.

A thin internal gateway — task type in; model, effort, speed tier and fallback out — keeps this in one place. Our generative and agentic AI team builds exactly this layer for clients, and the same pattern applies to the background agents many teams are now deploying.

What this means for teams building now

Do not rush to swap your default model to the newest flagship. Re-run your evaluation set on the new mid tiers first: if Opus 5.5, Sonnet 5.5 or GPT-6.1 Sol meet your bar, you can cut costs and latency at once. Then restructure prompts for caching, which pays off regardless of vendor.

1. 01This week: log cached tokens and cost per task.
2. 02Next: re-order prompts so the static prefix is cacheable; compare hit rates before and after.
3. 03Then: run your evaluation set on two or three mid-tier models at two effort levels and update your routing map.
4. 04In parallel: move model ids, effort and speed tiers into configuration, and wire a fallback provider.
5. 05Finally: keep Argon and other gated models on a watch list until they are generally callable.

If you would rather run some of this on your own hardware, our guide to self-hosted open-weight models covers when that beats an API. The short version: for most teams in Q4 2026, the biggest saving is not a new model but better caching and routing on the models they already use.

Frequently asked questions

Sources

1. Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic, 2026-09-01
2. Coverage of Claude Fable 5.1 and Mythos 5.1 pricing — VentureBeat, 2026-09-01
3. Claude Opus 5.5 — Anthropic, 2026-09-22
4. Claude Sonnet 5.5 — Anthropic, 2026-09-28
5. Anthropic newsroom — Anthropic
6. GPT-6 Astra (encyclopedia entry) — Wikipedia
7. Coverage of the GPT-6 Sol and GPT-6 Luna launch — MacRumors, 2026-09-22
8. Coverage of OpenAI DevDay 2026 — InfoQ, 2026-10-02
9. Coverage of OpenAI DevDay 2026 announcements — BGR
10. Gemini 4 Argon — Google, 2026-09-30
11. Coverage of the Gemini 4 Argon announcement — 9to5Google, 2026-09-30
12. Google AI updates, September 2026 — Google, 2026-10-02