Anthropic expands Claude's context window from 9K to 100K tokens, or ~75K words it can digest and analyze; OpenAI's GPT-4 has a context window of ~32K tokens

TechCrunch 2023-05-13 Kyle Wiggers

Context & Ripple Effects

Claude’s move put a 100K-token context window at the center of its comparison with GPT-4’s roughly 32K-token window, making long-document analysis a concrete competitive differentiator rather than a model-spec footnote.

The increase became a stepping stone in Claude’s product arc: Claude 2.1 later raised the limit to 200K tokens, and Projects applied that larger context to persistent workspaces for documents, chats, and code.

First-order effects

Claude can take in and analyze far larger bodies of text in a single interaction, reducing the need to split long material across multiple prompts.
Anthropic gains a clear specification advantage over the GPT-4 context-window figure cited in the coverage, while users can evaluate Claude for long-input workloads.

Second-order effects

OpenAI and other model providers face pressure to compete not only on output quality but also on how much user material their systems can retain in one request.
Long-document workflows become a more plausible target for AI products, increasing the value of prompt and document-management tools that fit useful material within a model’s available context.

Third-order effects

If providers continue expanding context limits, context capacity will become a core platform dimension alongside model quality, latency, and price—an arc reflected in Claude’s later 200K-context model pricing.
The differentiation may shift from raw window size toward context engineering: which information a product selects, preserves, and makes usable across ongoing work.

The trend: Foundation-model competition is moving toward larger, better-managed context windows that let AI systems work over fuller sets of user information.

Discussion

@anthropicai @anthropicai on x
Introducing 100K Context Windows! We've expanded Claude's context window to 100,000 tokens of text, corresponding to around 75K words. Submit hundreds of pages of materials for Claude to digest and analyze. Conversations with Claude can go on for hours or days. https://twitter.co…
@sarthakgh Sar Haribhakti on x
Anthropic (founded by OpenAI alums) expands Claude's context window (aka memory state) from 9K to 100K tokens or 75K words. Only available to biz customers via API. OpenAI's GPT-4 has a context window of ~32K tokens. https://www.theverge.com/... https://twitter.com/...
@karinanguyen_ Karina Nguyen on x
100k Tokens Context ~ the whole Great Gatsby book (75k words) ~ 6 hrs of audio content. You can treat as long-term memory and a way of finetuning! Some cool use cases: 1/ Put the whole API developer doc and ask complex reasoning tasks like prototype an app (e.g... https://twitter…
@aibreakfast @aibreakfast on x
Anthropic AI's new 100k token context window will allow for prompts equivalent to 250+ pages of text. This massive jump in token context length could allow students to upload entire textbooks for conversational analysis. https://twitter.com/...
@mathemagic1an Jay Hack on x
Retrieval-augmented generation isn't going away It still matters when: - huge amount of data - latency/cost is a major concern 100k tokens raised the high water mark on task that don't need fancy retrieval, however Projects predicated on fancy retrieval: take note https://twitter…
@anthropicai @anthropicai on x
We've made this available to our business partners, and are excited to see what they build. Read more here: https://www.anthropic.com/...
@_akhaliq @_akhaliq on x
Anthropic expandes Claude's context window to 100,000 tokens of text, corresponding to around 75K words blog: https://www.anthropic.com/... https://twitter.com/...
@mattshumer_ Matt Shumer on x
Huge step forward for LLMs. From my testing, Claude is quite intelligent. Now with 100,000 tokens, this will open up new use-cases that no other language model can currently match. https://twitter.com/...
@beyang @beyang on x
This is why LLM portability matters https://www.anthropic.com/.... If you're using Copilot, you have a 2-year old model with 2k tokens of context that doesn't know anything past 2021. If you're using Cody, you can use Claude, GPT-4, and the latest, greatest LLMs as they come onli…
@jonst0kes @jonst0kes on x
Anthropic's Claude drops a 100K-token context window. https://www.anthropic.com/... https://twitter.com/...
@criticalai @criticalai on x
Is this really worth the expenditure of water and energy? Asking w/ a desire to learn. Why would we want to do this? What would one be fine tuning _for_? & what would it's value be to a) the user b) technological progress c) society d) the planet? https://twitter.com/...
@bowtiedrobin @bowtiedrobin on x
Insane. Double checked all of the numbers. No hallucinations. But the wording on the “nearly $5.2 bil” is misleading, Netflix has $5.147 bil in cash and cash equivalents. Really great though. https://twitter.com/...
@patrickjblum Patrick Blumenthal on x
“Eliezer, it's Sam... Anthropic's 100K context window is dangerous. It feels very... unaligned. Something needs to be done about it. I'm sending you the GPS coordinates of their clusters now. Good luck.” https://twitter.com/... https://twitter.com/...
@emostaque Emad on x
LMs are like really talented grads that occasionally go off their meds that are really good at following instructions. Now the instructions are going to go to hundreds of thousands and then millions of words. This has quite a large impact on so many things.. https://twitter.com/.…
@omarsar0 Elvis on x
Claude's context window has expanded to 100,000 tokens of text! Context window is a big limitation of LLMs. We are about to see some mind-blowing stuff! https://twitter.com/...
@goodside Riley Goodside on x
100k tokens. Incredible. Demos all show long input; that's just low-hanging fruit. Think novel-length instruction — entire employee manuals as prefixes. Think k-shots — embed, take 200 nearest HDBSCAN clusters, sample 3 from each. “Long prompting” changes everything. https://twit…
@jerryjliu0 Jerry Liu on x
This adds another option in @karpathy's discussion of fine-tuning vs. retrieval. If context windows are big enough, should we always try to “stuff” the input prompt with context (e.g. an entire book)? What are the cost/latency tradeoffs? Excited to play around with this more. htt…
@mlpowered Emmanuel Ameisen on x
I'm very excited for this release. If you've used LLMs, you know how transformative a longer context can be. Feed it the complete docs for a library and have it use it Ask it detailed questions about every chapter of a textbook Lots you can do with 150-200 pages https://twitter.c…
@itsandrewgao Andrew Kean Gao on x
Anthropic's Claude AI blows past #GPT4 with a new 100,000 token context window (75,000 words). Enough to fit Harry Potter Book 1! Unsure how quality degrades (if it does) over length of convo. But if context windows can infinitely scale, embeddings may be in a little trouble.... …
@alexgraveley Alex Graveley on x
100k tokens is long enough to be in a relationship. https://twitter.com/...
@azeem Azeem Azhar on x
This has real applications. https://twitter.com/...
@kevinafischer Kevin Fischer on x
What happens to vector database companies here? Or langchain chat your data framework? https://twitter.com/...
@matthewclifford Matt Clifford on x
Wow. Super impressive. https://twitter.com/...
@abacaj Anton on x
Well this is definitely one way to surpass GPT-4. It will be interesting to see how it plays out on Claude. Hard to say how well their model will process so much text and what to expect from the latency Either way this is a massive context increase in such a short time https://tw…
@wintonark Brett Winton on x
4 years ago you could only prompt a language model with a few paragraphs Now you can feed one an entire book “Short term memory” for AI models has turned vertical. https://twitter.com/... https://twitter.com/...
@bentossell Ben Tossell on x
Context windows will just be a non-issue soon. Love it. https://twitter.com/...
@seconds_0 @seconds_0 on x
Holy cannoli, 100k token context window is incredible. Thats multiple whole papers, some whole code bases, entire 250 page novels https://twitter.com/...
@gestaltu Adam Butler on x
Haven't had a chance to test Claude, but a 100,000 token context window can manage most summarize/synthesize use cases without external handlers. In concert with handlers, massive private unstructured and structured document repos are open for complex queries. 🔥 https://twitter.c…
@nonmayorpete Pete on x
Anthropic announcing 100k-token context windows. GPT-4 has a 32k window in beta. Let's shoot for the moon, folks. 250K? 500K? 1M? Race is on. https://twitter.com/...
@simonw Simon Willison on x
Bah, only available to “business partners” the moment I'd love to know what the pricing is https://twitter.com/...

Chronicles