What's actually possible this quarter.
Every playbook in The Lens has to be grounded in what AI can really do today, not what the keynote demo promised. This page is the honest read, by capability, with concrete examples of what holds up and what falls apart. We update it every quarter as the models shift.
The model landscape this quarter
September was the busiest month of the year for frontier releases. OpenAI shipped GPT-6 Astra on 3 September, and Anthropic followed with Claude Opus 5.5 on the 22nd. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, and it costs $4/$20 per million tokens. Opus 5.5 now tops the Artificial Analysis Intelligence Index at 58, ahead of Fable 5.1 and GPT-6 Astra on 53. The index has been rescaled since June, so don't compare today's numbers with last quarter's. Google's Gemini 3.5 Pro, promised at I/O for June, still hasn't shipped.
The practical point for marketing teams is cost, not the leaderboard. Frontier-level output is getting cheaper by the quarter, and the mid-tier models (Claude Sonnet, GPT mini, Gemini Flash) now handle most drafting and every scoring gate in The Lens. Keep the expensive model for the jobs where judgement matters, such as positioning, synthesis and hero copy, and put the volume through the cheaper tier.
June made the case for staying model-agnostic. Anthropic's Fable 5 launched on the 9th, was suspended three days later under a US export-control order, and came back on 1 July once the order was withdrawn. Any team that had wired a pipeline to one named model spent that fortnight rebuilding. The Lens playbooks now describe models by tier (frontier or mid-tier) rather than by version, and list the models each workflow was tested on. The eval gates decide which model ships the work. In September we re-ran every prompt in every playbook on Claude Opus 5.5 against each playbook's own gates, fixed what broke, and published the test logs. GPT and Gemini weren't part of that run.
What changed for marketers since June
- ChatGPT is now a UK ad channel. Ads went live in the UK on 6 June, the first European market, shown to Free and Go users underneath answers, and UK advertisers now have an Ads Manager beta. Google's ads in AI Overviews still aren't running in the UK.
- Being cited in AI answers is its own game. The overlap between Google's top ten and the sources AI answers cite has fallen below 20%. That shift is in The citation moved off the page, and the seo-cluster-generator playbook is built for it.
- Sora is gone. OpenAI closed the app in April and switches off the API today, 24 September. Any video pipeline still calling it needs a new provider. Treat that as the argument for keeping generation behind a swappable step.
- Agents are moving into the apps you already use. OpenAI folded its Atlas browser into ChatGPT, and Claude in Chrome now runs as part of Cowork. Browser agents are becoming a feature of the main assistants rather than a separate product, which is how The Lens productivity stack assumes you'll use them.
- EU transparency rules now apply. The AI Act's Article 50 obligations took effect on 2 August 2026. AI providers must mark what their models generate. Brands using them must tell people when they're talking to an AI, and label realistic AI-generated or manipulated images, audio and video that could pass as real. If you market into the EU, your chat assistants, synthetic voice and deepfake-adjacent creative are in scope. In the UK, the ASA's position is technology-neutral: say you used AI where leaving it out would mislead.
The ratings below are the ones set in June. The quarter's releases could move two of them, generic video and branded-product imagery, and we've said below what we're watching. A rating only changes once a model has passed the relevant eval on live work.
The grading
Each capability gets one of four ratings.
- Production-ready, used on live client work without supervision beyond standard QA. The output ships.
- Production-with-guardrails, works in the pipeline but needs an eval gate, a fact-check pass, or a human spot-check before it ships.
- Experimental, useful for internal drafting or as a starting point a human heavily edits. Not yet shippable as the final artefact.
- Not yet, the demo looks great. Real-world reliability is too low for production work. Don't promise it to a client.
Text generation and structured analysis
This is where AI compounds time hardest. Drafting a long-form piece from a structured brief, extracting a brand voice from a corpus, classifying 500 search queries by intent, surfacing contradictions across a brand's public surface, all of this works reliably with the right scaffolding. The Lens's drafting, voice, positioning and SEO-cluster playbooks all sit in this category.
Caveat. Factual claims need grounding. The model will assert plausible- sounding numbers and quotes if you don't constrain it. Every production drafting pipeline in The Lens runs a fact-grounded eval gate that verifies claims against approved sources.
Voice and tone matching
A draft scored against a 12-check voice rubric ships. A draft based on "write in our voice" without a profile produces generic copy every time. The brand-voice-extraction playbook builds the observable profile, and everything downstream relies on it.
Original research and fact-finding
Race results, athlete bios, sponsorship deal terms, regulatory changes, these need real sources retrieved at run-time, not recall from training. The earned-media-pitch playbook hard-fails any reference that wasn't fetched and verified, precisely because a hallucinated journalist quote burns relationships permanently.
Image generation, generic subjects
Generating a wide shot of a road cyclist at dawn on a coastal switchback, a trail runner cresting a ridge, a swimmer at lap 60 of a session, this works. The output is usable for social backdrops, brand mood pieces, blog hero images, paid ad creative variants. Eval gate is human review for hand quirks, bike geometry, and gear plausibility.
Image generation, branded product
We've tested every frontier model on this every quarter for two years. None produce reliable brand-accurate product imagery. Wheelsets get the wrong spoke count. Logos warp. Decals end up in the wrong place. The shoe outsole pattern is generic. The kit's chevrons go the wrong way. For product hero shots, you still need a real shoot. AI is for the b-roll and the moodboard, not the product card.
Workaround that does work is to shoot the product once, then use image-to-image with the real product as the source to produce variants in different environments (different light, different terrain, different season). The model edits the surroundings, not the product. Eval gate is to pixel-diff the product region against the source.
Watch this quarter. The newer image-editing models (Google's Nano Banana 2, OpenAI's GPT Image 2) are sold on keeping the source image consistent through an edit, which is exactly what the workaround above needs. Until one clears the pixel-diff gate on live product work, the rating stays where it is.
Video generation, generic subjects
A wide of a cyclist on a Tuscany hill at golden hour, a runner climbing wooden stairs in a forest, a swimmer's underwater push-off from a wall, these work. Useful for social cuts, paid creative variants, atmospheric brand films where named athletes and named products aren't required. Eval gate is human review for limb physics, water dynamics, and the recurring "wrong number of pedal rotations per crank stroke" issue.
Watch this quarter. With Sora retired, the realistic options are Google's Gemini Omni (which replaced Veo in the Gemini app and added edit-by-conversation), Kling and Runway. Build the video step so the provider can be swapped without rewriting the pipeline around it.
Video generation, branded product or named athlete
Generating video of your specific bike being ridden through an environment, with logos intact through every frame, with accurate geometry maintained through motion, doesn't work today. Logos warp. The bike morphs across frames. The kit colour drifts. Generating video of a named athlete who hasn't been licensed for AI use is legally fraught even where the model is convincing.
Workaround that does work is to shoot the athlete and the product practically. Use AI for the supplementary cuts that don't need the named athlete or product in frame, such as environmental shots, time-of-day variants, weather-condition variants for the same campaign.
Audio generation, synthetic voice
Quality is good enough for production in most contexts. The ethical gate matters more than the technical one, because listeners feel misled if a brand pretends a synthetic voice is the founder's actual voice. Be honest. The brands that disclose are winning trust, and the brands that don't are losing it when found out.
Audio generation, music
15 to 60 second pieces for social cuts and reels work. Full-length original tracks for hero films still struggle with structural coherence past the 90-second mark. Licensing is also evolving, since the legal landscape for AI music in commercial use is changing every few months, and we re-check this section every quarter.
Browser automation and research
Pulling a journalist's last five articles, scraping a competitor's product range, gathering GSC data, classifying a corpus of reviews, all production-ready. The browser-using agents have gotten reliable enough that this is now a workhorse capability in most Lens playbooks.
Data analysis and visualisation
Channel-mix simulators, attribution tear-downs, retention cohort analyses, all reliable with structured data input. The model writes the analysis script (Python, R, SQL), runs it, interprets the result. Just don't trust visualisations the model invents from prose, since the chart should come from the data, not from the description.
The honest summary for endurance brands
Most of the leverage is in analysis, drafting and orchestration. The model compounds time across positioning, voice, content production, channel-mix decisions, lifecycle journeys, race-result recaps, and the long tail of operational work that fills a marketing team's week.
The shoot still matters. Athletes, products, named team imagery, these are practical work. You still hire a photographer. You still fly to Tuscany. You still pay the athletes. AI augments the shoot's output (variants, social cuts, environmental b-roll) but doesn't replace the shoot itself.
Brands that pretend otherwise, with claims like "we used AI for our whole campaign, including the product hero", are either lying or shipping work that's hurting the brand. The brands actually winning treat AI as the most productive intern they've ever had, brilliant at the groundwork, supervised on the edge cases, and never trusted with the family silver until the eval gate proves they should be.
How this page evolves
We re-grade every capability quarterly. When a model lands that moves a rating, we annotate the entry with the date and the evidence. When a capability is downgraded, and yes, it happens, we explain what stopped working and what the workaround is.
Subscribers get the change log, and everyone gets the current state.
Honest scope. Real work.
Every playbook in the library is grounded in what's on this page. When the capability landscape shifts, the playbooks shift with it.