The knowledge cutoff

A model's knowledge has a recency boundary, and the boundary that vendors document is graded rather than clean.

Anthropic's model table gives two dates for each model. One is the reliable knowledge cutoff, described as "the date through which the model's knowledge is most extensive and reliable", and the other is a separate, broader training data cutoff.¹ Claude Sonnet 5, Opus 5 and Fable 5.1 carry the same date for both; Claude Haiku 4.5 does not, with a reliable cutoff of February 2025 against a training data cutoff of July 2025.¹ Google's card for Gemini 3.5 Flash-Lite states March 2026 and then warns that coverage is uneven across domains, with some areas still limited to January 2025.² So the published date is exact, and what it marks is the end of the period through which the model's knowledge is most extensive and reliable, not a wall with knowledge behind it and emptiness in front. The figure below draws that shape in order: uneven coverage before the reliable cutoff, a further span of possible training data running on to a separate, broader training-data cutoff, and only past that a true edge, with the model's release day marked apart from both.

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 700 330" font-family="system-ui, sans-serif" fill="currentColor">
<title>A model's knowledge over time: uneven coverage by topic, a reliable cutoff, a further span of possible training data to a separate training-data cutoff, and a later release day</title>
<text x="60" y="34" font-size="17">What the model was trained on, over time</text>
<text x="60" y="58" font-size="15" opacity="0.75">Coverage by topic is uneven even before the reliable cutoff (bar heights illustrative)</text>
<g opacity="0.55">
<rect x="70" y="150" width="20" height="60" />
<rect x="100" y="162" width="20" height="48" />
<rect x="130" y="138" width="20" height="72" />
<rect x="160" y="170" width="20" height="40" />
<rect x="190" y="144" width="20" height="66" />
<rect x="220" y="180" width="20" height="30" />
<rect x="250" y="132" width="20" height="78" />
<rect x="280" y="158" width="20" height="52" />
<rect x="310" y="166" width="20" height="44" />
<rect x="340" y="142" width="20" height="68" />
<rect x="370" y="174" width="20" height="36" />
</g>
<line x1="60" y1="210" x2="685" y2="210" stroke="currentColor" stroke-width="1.5" />
<line x1="400" y1="86" x2="400" y2="224" stroke="currentColor" stroke-width="2.5" />
<text x="395" y="78" font-size="15" text-anchor="end">Reliable cutoff</text>
<rect x="400" y="196" width="70" height="14" opacity="0.22" />
<text x="435" y="250" font-size="14" text-anchor="middle" opacity="0.85">possible training</text>
<text x="435" y="268" font-size="14" text-anchor="middle" opacity="0.85">data continues</text>
<line x1="470" y1="86" x2="470" y2="224" stroke="currentColor" stroke-width="2" stroke-dasharray="5,4" />
<text x="475" y="78" font-size="15" text-anchor="start">Training data cutoff</text>
<text x="580" y="140" font-size="15" text-anchor="middle">Nothing trained in</text>
<text x="580" y="160" font-size="15" text-anchor="middle">beyond this cutoff.</text>
<text x="580" y="180" font-size="15" text-anchor="middle">Current material can</text>
<text x="580" y="198" font-size="15" text-anchor="middle">be supplied in chat.</text>
<circle cx="650" cy="210" r="5" />
<text x="650" y="240" font-size="15" text-anchor="middle">release day</text>
<text x="60" y="300" font-size="15" opacity="0.75">The reliable cutoff and the training-data cutoff are different documented dates.</text>
</svg>

Figure: the reliable cutoff is not a wall; training data may continue to a separate, later cutoff before the true edge and release day.

Two further vendor documents say the same thing in a different way. Claude Opus 5 carries a reliable cutoff of May 2026 while its own system card is dated 24 July 2026, so a model's knowledge ends before its release.³ OpenAI's page for GPT-5.5 gives a December 2025 cutoff with a default snapshot from April 2026.⁴ Amazon's card for Claude Opus 4.8 lists a single knowledge cutoff of January 2026, which is the one-date shape most readers already expect.⁵

What this looks like in use: a model whose documented training-data cutoff predates last month is not short of a small update on those events, it was never trained on that month at all. Pasting the material into the chat is a different mechanism, and one route by which current material can reach the model. Anthropic's documentation separates the two explicitly, describing the context window as working memory and distinct from the corpus the model was trained on.⁶

The judgement stops at the edges of what is documented. These dates are per model and per vendor and they move with every release, so a date you read once is not a property of the product. One vendor publishing two dates is not evidence that another vendor's single date conceals a second. And the uneven-by-topic warning is documented for one model family, not established as a general rule.