A story has been making the rounds: Anthropic supposedly launched “Claude Haiku 4.5” on March 6 with native vision capabilities, slashed input pricing to $0.80 per million tokens, and hit 70% of “Opus 4.6” accuracy on document processing tasks. It sounds like a solid enterprise AI scoop. The problem is that none of it checks out.
Promptyze ran the claims against Anthropic’s official model documentation, pricing pages, and newsroom — and found no matching announcement, no such model version, and pricing figures that don’t line up with anything Anthropic has published. Here’s what the actual record shows.
Anthropic’s current Claude lineup uses version 3: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus. There is no publicly released “Claude Haiku 4.5” — and no “Opus 4.6” either, which makes the claimed accuracy benchmark impossible to verify by definition. Anthropic has not announced a Claude 4 generation as of March 2026 in the way described in the original brief.
Claude 3 Haiku does have vision capabilities — that part is true. Anthropic built multimodal support into the entire Claude 3 family at launch in March 2024. But that’s not a March 2026 news event. It’s a two-year-old feature.
The $0.80 per million input tokens figure doesn’t match Anthropic’s published rates for any current Claude model. Claude 3 Haiku is positioned as the most affordable model in the lineup, but the specific number cited in the brief contradicts official pricing documentation. No tech outlet — TechCrunch, VentureBeat, The Verge, none of them — reported a March 6 pricing change or model launch matching these details.
Where did the story originate? Hard to say with certainty. The most likely candidates: an AI-generated summary that hallucinated product details, a speculative post that got laundered into a “report,” or a misread of a roadmap discussion. Whatever the source, the claims arrived without a paper trail.
For anyone who came here because they genuinely want cheap, fast document processing with Claude, the real picture is still useful. Claude 3 Haiku handles images, PDFs, and screenshots natively through the Messages API. It’s the fastest and cheapest Claude 3 model, designed for high-volume, latency-sensitive workloads. Batch API support across Claude models is real and does reduce costs meaningfully for asynchronous document pipelines.
If you’re building a document processing workflow today, Claude 3 Haiku with vision is a legitimate option worth benchmarking. Just base your architecture on what Anthropic has actually shipped and priced — not on a spec sheet that doesn’t exist.
Fabricated or hallucinated AI product announcements are becoming a genuine problem. When a model confidently generates a fake press release — complete with pricing, dates, and accuracy benchmarks — and that output gets passed upstream as a source, the damage spreads fast. Publishers run corrections, engineers make architectural decisions based on phantom specs, and trust erodes a little more each time.
Promptyze fact-checks briefs before publishing. In this case, the research stage caught a story that had zero verifiable foundation. We’re publishing the correction instead of the fiction — because that’s what the record deserves.
