Every major frontier model now claims a 1M token context window. That claim is technically accurate: the models accept that many tokens. Whether they reliably retrieve information from the middle of those contexts is a different question, and the answer varies far more than the spec sheets suggest.
The most useful benchmark for this is MRCR v2, which tests retrieval recall at multiple context lengths: 32K, 128K, 512K, and 1M. Unlike toy needle-in-a-haystack tests, MRCR v2 requires the model to locate and synthesize information across realistic document structures. The scores below are drawn from BenchLM.ai's June 2026 leaderboard.
The 1M-context lineup in 2026
| Model | Context | MRCR v2 @ 128K | MRCR v2 @ 1M | Input $/M | Output $/M |
|---|---|---|---|---|---|
| GPT-5.5 | 1M | 94.8% | 74.0% | $5.00 | $30.00 |
| Claude Opus 4.8 | 1M | ~85% (est.) | Not published | $5.00 | $25.00 |
| Gemini 3.1 Pro | 1M | ~90% (est.) | 26.3% | $2.00–$4.00 | $12.00–$18.00 |
| Gemini 3.5 Flash | 1M | 77.3% | 26.6% | $0.30 | $2.50 |
| Llama 4 Scout | 10M (theoretical) | Strong to ~32K | Degrades sharply | Free (self-hosted) | N/A |
| DeepSeek V4 | 1M | Competitive | Not published | $0.14 | $0.28 |
Claude Fable 5 (1M context, $10/$50 per 1M tokens) is excluded from the active ranking: Anthropic suspended access following a US government export-control directive on June 12, 2026. Claude Opus 4.8 is the current Anthropic option.
Pricing verified June 2026. Gemini 3.1 Pro's tiered pricing applies above 200K-token prompts (input doubles from $2 to $4; output from $12 to $18).
Why the advertised number misleads
Accepting 1M tokens and reliably retrieving specific information from 1M tokens are not the same thing. The "lost in the middle" effect is consistent across frontier models: information near the start and end of a long context is recalled significantly better than information buried in the middle.
Research from LMU Munich and Adobe Research, documented in the NoLiMa benchmark, found that 11 of 13 tested LLMs dropped below 50% of their short-context baseline scores at 32K tokens when the task required multi-hop reasoning rather than surface-level pattern matching. The models got there far sooner than their advertised maximums would imply.
Our context window explainer covers what different window sizes actually hold and where quality drop-off typically kicks in. The practical upshot: a well-structured 32K prompt usually outperforms a sprawling 800K one. The use cases where the full 1M window earns its cost are specific: whole-codebase analysis where chunking loses cross-references, legal document review spanning hundreds of pages, or long-running agent workflows where context accumulates over many turns.
The models, in practice
GPT-5.5 is the standout for genuine long-context work. Its MRCR v2 score drops from 94.8% at 128K to 74.0% at 1M, a real fall, but still far above every other published score at that length. If retrieval accuracy across a fully-loaded context is the requirement, GPT-5.5 is the only frontier model with published evidence that it holds up. Pricing is $5/$30 per million tokens with no long-context surcharge. The DataCamp comparison puts GPT-5.5's long-context lead in context against current Gemini variants.
Claude Opus 4.8 is the best available Anthropic option at 1M context. Anthropic doesn't publish MRCR scores, but our benchmark comparing Opus 4.8 to GPT-5.5 found it competitive on document-heavy reasoning tasks at 12 prompts across reasoning, extraction, and analysis. The key structural advantage: Anthropic charges the same per-token rate whether the prompt is 9K or 900K tokens, with no penalty for pushing the window.
Gemini 3.1 Pro and 3.5 Flash share a striking result: both score around 26% on MRCR v2 at 1M, virtually identical despite Pro being significantly more expensive. That number means roughly one in four pieces of information buried mid-context surfaces reliably. For tasks that actually rely on retrieving specific content from long documents, this is a meaningful constraint. Gemini Flash remains excellent for classification, routing, and short-output extraction where the relevant signal is near the document boundaries. Its 1M context is more useful as a buffer than as a full working space.
Gemini 3.1 Pro also carries a pricing surcharge for prompts above 200K tokens: input doubles from $2 to $4 per million, output from $12 to $18. That surcharge matters more than the base rate for teams with frequent long-context workloads.
Llama 4 Scout claims 10M tokens, which exceeds every other model by at least 10x. The practical ceiling is much lower. Independent testing on DEV Community found effective context degrading around 32K tokens on standard hardware, with first-token latency exceeding 60 seconds at the full 10M window. Reaching that ceiling requires 8 H100 GPUs, a $60K hardware floor. Llama 4 Maverick's 1M window is more accessible; Scout's 10M is a headline for use cases involving extremely large codebase or document sets on high-end infrastructure. For teams already running large GPU clusters, it's real. For anyone else, Maverick is the better pick. The best open-source LLMs for on-prem post has full hardware requirements for running Llama 4 variants locally.
DeepSeek V4 at $0.14/$0.28 per million tokens is the cheapest path to a 1M context window. No published MRCR data at 1M, but our RAG synthesis benchmark found it competitive on long-document answer synthesis tasks. For teams where 1M context is useful-to-have rather than essential, and where cost is the primary constraint, DeepSeek V4 is worth testing first.
What it costs to fill the window
Filling a full 1M-token prompt once:
| Model | Cost for 1M-token input | Long-context surcharge |
|---|---|---|
| DeepSeek V4 | $0.14 | None |
| Gemini 3.5 Flash | $0.30 | None |
| Gemini 3.1 Pro (>200K) | $4.00 | 2x above 200K |
| Claude Opus 4.8 | $5.00 | None |
| GPT-5.5 | $5.00 | None |
A 1M-token prompt holds roughly 750,000 words, equivalent to a large codebase, three to four full novels, or fifty research papers. For workloads that repeatedly pass the same base document with different queries, prompt caching cuts this cost substantially. Anthropic caches at 90% discount after the first pass; OpenAI at 50%.
Subscription vs API for long-context users:
| Provider | API price (input/output per 1M) | Consumer subscription | Notes |
|---|---|---|---|
| OpenAI (GPT-5.5) | $5 / $30 | ChatGPT Plus $20/mo | No long-context surcharge |
| Anthropic (Opus 4.8) | $5 / $25 | Claude Pro $20/mo | No long-context surcharge |
| Google (Gemini 3.1 Pro) | $2/$12 base, $4/$18 >200K | Google One AI Premium $20/mo | Surcharge makes API costlier for frequent long-context use |
Break-even for Claude Pro vs API: at $5/M input, you hit $20/month after roughly 4 million input tokens, or around 30 full 128K-token requests per day. Below that volume, the subscription covers long-context use at flat cost. Google's surcharge changes the math significantly for Gemini users: if your prompts regularly exceed 200K tokens, the $20/mo Google One subscription often beats the API rate.
Pricing verified June 2026: OpenAI API pricing, Anthropic pricing, Google AI pricing.
Verdict
If you need reliable retrieval from 1M+ token contexts, GPT-5.5 is the only model with published evidence supporting that claim (74% MRCR at 1M). Claude Opus 4.8 is the strongest alternative without a published score: no surcharge, competitive quality on reasoning-heavy document tasks, and available now.
Gemini models offer the largest context windows for the cheapest base rate, but their 26% MRCR at 1M makes them poorly suited for dense mid-context retrieval. Use them when you need a big buffer for long conversations or documents where you'll mostly read the endpoints, not when you need to find a specific clause on page 200 of a contract.
For teams routing across providers based on context length or task type, the LLMTest proxy handles model selection at the API layer, routing short prompts to cheaper models and escalating to GPT-5.5 or Opus 4.8 only when the full window is needed.