Claude Opus 5.5 and GPT-6 Sol for SEO: What the New Models Change for AI Search Visibility

On 22 September 2026, Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna, about ninety minutes apart. The model coverage, benchmarks and the head-to-head comparison are in separate articles.

This one answers a narrower question: what do these two launches change for being found, used and cited by AI? The general groundwork, such as agent traffic in analytics, the friction audit and content structure, is covered in what GPT-6 Astra changes for SEO and still applies. What follows is what's new this week.

1. Why these two launches matter more for search than the flagships

Flagship models set records. Mid-tier models set volume. Most AI answers people actually read, whether in a support widget, a coding agent, a browser extension or a research tool, come from whichever model is good enough and cheap enough to run at scale. This week that tier got much cheaper:

Prices and availability: Confirmed. When a capable model costs a tenth of a cent per thousand words, developers put AI answers into places that never had them: product pages, internal search, email triage, comparison tools. Each of those is a new place where your content either gets retrieved and credited, or doesn't. Analysis

There's also a quieter shift. Both launches promise plainer, more direct answers: Anthropic says Opus 5.5 "puts the most important information up front", and OpenAI promises "less jargon" and "slightly shorter answers". Confirmed Shorter answers mean fewer sources quoted per answer, so the competition for each citation gets tighter. Speculation

2. The knowledge windows: where each model has to look things up

A model's knowledge cutoff marks where memory ends and retrieval begins. Here's the current map, from each vendor's documentation:

That gives three zones to plan content around:

1. Before about February 2026: trained territory. Every current model has seen it. You're competing with what the model already "believes", and changing that takes repeated, consistent coverage across many sources over time.
2. April to June 2026: the split zone. Opus 5.5 and Fable 5.1 may know it from training; GPT-6 Sol, Astra and regular ChatGPT have to retrieve it. The same question can get a memorised answer from Claude and a retrieved, cited answer from ChatGPT.
3. July 2026 onwards: open territory. Every model has to search. There's no trained opinion to overcome, so a well-structured, clearly dated page can become the answer quickly.

These zones are an inference from the published cutoffs. A cutoff isn't a hard wall, and coverage thins out in the final months before it. Analysis

3. Six crawlers, two companies, separate switches

Neither model can cite a page its company's crawlers can't reach. Both companies now split crawling by purpose, and each switch works independently:

From OpenAI's crawler documentation and Anthropic's help centre. Confirmed OpenAI also runs OAI-AdsBot to check ad landing pages, and says robots.txt changes take about 24 hours to affect search. Anthropic says all three of its bots honour robots.txt, including Claude-User, and that blocking ClaudeBot does not block the other two. Confirmed

If your aim is to be cited but not used for training, the robots.txt logic looks like this:

# Visible in AI search and user-initiated fetches
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: ChatGPT-User
Allow: /

User-agent: Claude-User
Allow: /

# Training is a separate, content-rights decision
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

Whether to allow training is a business decision, not an SEO one. The llms.txt and AI crawler guide covers the trade-offs. The common mistake is different: a robots.txt written in 2024 that names GPTBot and ClaudeBot but never mentions the search and user agents, or a CDN bot-protection rule that silently blocks all six. Analysis

4. Search economics: why cheap models may search less

This is the least discussed effect of the price cuts, and possibly the most important one for visibility.

Both companies charge a flat fee for their built-in web search tool: $10 per 1,000 searches on the Claude API, and $10 per 1,000 calls listed for OpenAI's reasoning models, plus the tokens the results add. Confirmed That fee didn't fall this week. Token prices did.

Calculation at list input prices. On Luna, one search costs ten times as much as reading a whole page. A developer building a high-volume product on a budget model has a real reason to limit web search: answer from the model's memory where possible, and search only when a question clearly needs fresh facts. Speculation

If that happens at scale, it cuts both ways for SEO:

• Training-data presence matters again. Cheap assistants that rarely search answer from what they learned, so being well represented in the sources models train on (consistent facts, widely cited pages, a clear entity footprint) is back to being a visibility lever, not just a legacy one.
• Freshness signals become the trigger. If apps search only when a question looks time-sensitive, pages that make their dates and "as of" facts explicit are more likely to be what those searches find.

5. A cross-model citation test (about 90 minutes)

You don't need a tool to start. A spreadsheet and two browser tabs are enough.

1. Write 20 prompts your buyers would really ask: five "best X for Y", five "X vs Y", five "how much does X cost", and five about something that changed in 2026.
2. Tag each prompt by knowledge zone: trained (before February 2026), split (April to June 2026) or open (July 2026 onwards).
3. Run every prompt in Claude and in ChatGPT with web search on, in a fresh chat each time.
4. Log two separate columns: mentioned (your brand is named) and cited (your page is linked as a source). They're different outcomes with different fixes.
5. Log the competitors that are mentioned and cited in your place.
6. Read the pattern by zone. Missing in the open zone usually means crawling or structure. Missing in the trained zone means a weak footprint across sources. A split between Claude and ChatGPT in the split zone is often just timing.
7. Re-run monthly. Answers vary from run to run, so trends across months mean more than any single result.

A worked template for logging this is in Notion for SEO (the AI Answer Log system), and the scoring method is in the AI search visibility benchmark.

6. Content moves that work on both models

The fundamentals are covered in how to get cited by AI search. Four points matter particularly this week:

• Put an explicit date on every fact that changes. "As of 22 September 2026, GPT-6 Sol costs $2 per million input tokens" gives a model something it can lift and a date to judge freshness by. A sentence like "prices recently dropped" gives it nothing.
• Answer first, then explain. Both launches promise shorter, more direct answers, and a model writing a short answer takes the clearest single sentence it can find.
• Use tables for anything comparable. Specs, prices and pros and cons in a real HTML table are easier to extract accurately than the same facts spread across paragraphs.
• Label claims by how certain they are. Pages that separate confirmed facts from speculation give a model a safer quote, which matters more now that both vendors are optimising against hallucination and OpenAI's Sol answers fewer questions when unsure. Analysis

7. Which model for which SEO task, and what it costs

These models are also tools for SEO work. Estimated list-price costs for common jobs are below. The token counts are illustrative and exclude reasoning tokens, which both companies bill as output, so real costs at higher effort settings will be higher. Calculation

* Above 272K input tokens, OpenAI bills the whole request at 2× input and 1.5× output. That's why Sol costs almost exactly the same as Opus 5.5 on the 300K audit ($1.35 against $1.40), instead of about half. At that price, the stronger model is the better buy. Confirmed pricing rule.

The pattern: Luna for anything repetitive with a clear right answer, Sol for drafting, and Opus 5.5 for judgement and very long inputs. Opus 5.5 also has the newer knowledge cutoff, which matters for strategy work about 2026 search changes. At these prices the model bill is rarely the real cost. The review time for a wrong classification or a weak brief almost always is. Editorial view

You can run your own numbers in the Opus 5.5 vs GPT-6 Sol cost calculator.

8. A two-week plan

9. Frequently asked questions

Work with me

Be the source both assistants cite

I'm Jayant Solanki, an SEO, GEO and automation strategist working with eCommerce, local-service and global brands. Every new model moves the line between what AI remembers and what it looks up. I help brands be on the right side of it in Claude and ChatGPT alike.

A GEO engagement typically covers:

• Crawler access for all six AI user agents, checked against real logs
• Cross-model citation testing, with knowledge-zone analysis
• Content restructured so a model can lift a specific, dated claim
• AI-assisted SEO workflows priced to the cheapest model that does the job
• Measurement that honestly separates agent traffic from human traffic

Book a technical review → How AI picks sources

Ranked #1 for "metal buildings" · +30% YoY organic traffic · Evidence-labelled research

Sources

• Claude Platform Docs, Models overview (knowledge cutoffs) and Pricing (web search at $10 per 1,000 searches)
• OpenAI API docs, GPT-6 Sol and GPT-6 Luna model pages; Pricing (web search tool)
• OpenAI, Overview of OpenAI crawlers
• Anthropic Help Center, Does Anthropic crawl data from the web, and how can site owners block the crawler?
• Anthropic, Introducing Claude Opus 5.5; OpenAI, Introducing GPT-6 Sol and Luna
• OrcaRouter, Claude Opus 5.5 vs GPT-5.6 Sol (GPT-5.6 Sol knowledge cutoff)

Jayant Solanki

AI-Ready SEO, GEO & AIO strategist based in Indore, India, working with eCommerce, local-service and global brands across India, the UAE and the US. This page will be updated as GPT-6 Sol reaches regular ChatGPT and as the crawler documentation changes.

Google Analytics Certified (2026–2027) · verify

More about Jayant →