explainx / blog
Anthropic’s Jul 24 zoom/crop cookbook: Fable 5 Chartography 29%→73%, Sonnet 5 13%→44%. How 28×28 patches, full-res crops, and cost trade-offs work.
Jul 20, 2026
A July 19 X thread revived "common knowledge" that Claude Sonnet is 1T params, Opus 5T, Fable/Mythos 10T. explainx.ai maps Musk's corroborating Grok ratios, unexcitedneurons throughput estimates, Kimi K3 benchmark paradox, and why Anthropic still publishes zero official counts.
Jul 14, 2026
Claude Pro at ₹2,000/mo annual (₹2,399 monthly), Max from ₹11,999, Team from ₹2,399/seat — India's #2 Claude market finally bills in rupees. explainx.ai breaks down GST, payment gaps, and Fable 5 access.
Jul 12, 2026
Anthropic's Fable promo ends Sunday July 12 at 11:59pm PT. From July 13 every Fable token runs on usage credits at API rates. explainx.ai explains what changes today and how to cut costs with Sonnet + advisor patterns.
On July 24, 2026, Anthropic’s developer channel refreshed the multimodal zoom tool cookbook: when a chart label or line crossing is too small after downscaling, Claude can request a region and get a magnified crop from the full-resolution original.
That is not a new weights drop. It is the agent harness move for vision — give the model a lean-in tool instead of hoping a bigger Opus ships tomorrow. The day after Opus 5 failed to launch, ClaudeDevs shipped reading glasses.
| Question | Answer |
|---|---|
| What shipped? | Updated zoom / crop cookbook in anthropics/claude-cookbooks |
| Core idea? | Crop from full-res original, magnify into the image budget |
| Chartography (Fable 5)? | 29% → 73% (+44 pts) with tool |
| Chartography (Sonnet 5)? | 13% → 44% (+31 pts) with tool |
| Why it works? | 28×28 visual patches; tiny detail dies on downscale |
| Cost tax? | ~$0.08→$0.95 (Fable) · ~$0.02→$0.64 (Sonnet) per Q |
| New model? | No — tool loop on existing Fable 5 / Sonnet 5 |
| Demo sanity check? | Dense Widget-A vs Widget-E: 10% → 100% on 40 runs |
Claude cannot “lean in” on its own. It sees an image once, at a fixed effective resolution. When a peak annotation or an 8-pixel line gap sits below that resolution, prompting does not recover the pixels.
Anthropic’s vision stack measures images in visual tokens — one token per 28×28-pixel patch — under edge and token budgets. On the high-resolution tier used by claude-fable-5 and claude-sonnet-5, the cookbook cites up to 2576 pixels on a side and 4784 visual tokens. Images over budget are automatically downscaled before Claude sees them. Small details scale down with them.
That is why dense charts fail cold — and why a zoom tool that crops from the original (not the downscaled copy) is the fix.
The agentic loop is simple:
resized_size() so coordinates map 1:1).zoom(x1, y1, x2, y2) in absolute pixel coordinates.handle_zoom() maps those coords onto the full-resolution original, crops, and upscales the crop to fill the image budget (maximize patches on the detail).Design rules that matter in production:
| Rule | Why |
|---|---|
| Pre-resize client-side | Coords Claude emits match the image you hold |
| Crop from full-res original | Recover pixels lost to automatic downscale |
| Absolute pixels, not 0–1 | Matches Anthropic vision-coordinates guidance |
| JPEG for zoomed crops | Tool results accumulate; PNG crops can hit the 32 MB request limit |
| Long histories | Prefer the Files API so payloads stay small |
Charts are the running example. The same pattern applies to dense PDFs, UI screenshots, schematics, scanned forms, and label photos — anywhere detail is small relative to the frame. For Sonnet 5 day-to-day and Fable 5 credit burns, this is harness work, not a model wait.
Anthropic measured Chartography (Surge AI): 100 chart-reading questions over dense real-world charts. Every question asked twice per model (with / without tool); free-form answers graded with a model judge per the dataset card.
| Model | No tool | With zoom | Delta |
|---|---|---|---|
| claude-fable-5 | 29% | 73% | +44 pts |
| claude-sonnet-5 | 13% | 44% | +31 pts |
Accuracy run dated 2026-07-18 against the public Claude API. Mean per-question costs (30-question sample, 2026-07-19, judge excluded, Sonnet at intro price):
| Model | No tool | With zoom |
|---|---|---|
| Fable 5 | ~$0.08 | ~$0.95 |
| Sonnet 5 | ~$0.02 | ~$0.64 |
The tool buys accuracy with tokens: each zoom re-sends conversation context plus a magnified crop. Zoom arms used prompt caching; no-tool arms were single uncached calls. An optimized client that caches the wrap-up turn runs a bit cheaper than the published means.
Held-out caution: Chartography is an evaluation set — do not train on it (the dataset card includes a canary). Prompts/answers metadata are Surge AI, CC BY 4.0; chart images remain with original rights holders.
The cookbook’s self-generated dense chart asks: at month 48, which line is higher — Widget-A or Widget-E? The lines sit 26 units apart (8 pixels in the render), next to a real crossing — a fraction of one 28×28 patch.
| Condition | Result (recorded) |
|---|---|
| Cold (no tool) | Widget-A wrong in 36/40 (90%) |
| With zoom tool | Widget-E correct in 40/40 |
Claude zooms repeatedly, each crop tighter, then reads the ordering from the magnified view. That is geometry failure, not “the model is dumb” — and it matches what model vs effort debates miss when the bottleneck is pixels per patch, not thinking tokens.
crop_tool.ipynb in anthropics/claude-cookbooks (multimodal/). Schema + handle_zoom are designed to copy.max_iterations / stop-reason handling so thinking-heavy turns do not truncate mid-zoom.Anthropic’s sample costs show zoom turning an $0.08 Fable question into ~$0.95. That is fine for high-stakes chart QA and painful for bulk document sweeps. Practical controls:
| Control | Effect |
|---|---|
| Sonnet 5 first | Smaller base + still +31 Chartography points |
| Max zoom iterations | Cap runaway tighten-crop loops |
| Cache the conversation prefix | Cookbook notes zoom arms used caching; wrap-up can miss — fix client-side |
| JPEG crops + Files API | Stay under 32 MB; keep history lean |
| Pre-filter images | Don’t attach a 8K screenshot if the answer lives in one panel — crop client-side first |
| Advisor/executor split | Cheap model decides whether to zoom; spend Fable only when needed (pairing guide) |
Also remember: magnification helps legibility, not hallucination discipline. A wrong legend color with perfect zoom is still wrong — keep graders and golden answers for your domain.
resized_size() and zooms land on the wrong region.The Jul 24 zoom cookbook is Anthropic’s clearest public proof that vision failures are often resolution budgets, not model IQ. Give Claude a crop-from-original lean-in tool and Chartography roughly doubles-plus for Fable 5 and more than triples for Sonnet 5 — at a predictable token tax.
If you were waiting on a Thursday Opus drop to fix chart QA, ship the zoom loop instead.
Sources: Claude cookbooks — multimodal crop / zoom tool · ClaudeDevs announcement (Jul 24, 2026) · Chartography (Surge AI) · Anthropic vision / tool-use docs
Accuracy and cost figures reflect Anthropic’s cookbook measurements dated July 18–19, 2026 against the public Claude API. Image budgets, pricing, and cookbook code change — re-verify against the live notebook and platform docs before production.