Palmyra X5: Writer's 1M-Context Agent Model Before Its Sunset

Palmyra X5 has a lopsided design, and that is the useful thing to understand about it. It reads a million tokens cheaply and writes at most 8,192 of them, at ten times the input price. It is a model for reading large amounts of material and answering briefly, and a poor fit for almost anything else.

It also has a fixed end date. WRITER's model documentation schedules Palmyra X5 for deprecation on 14 December 2026 and recommends Palmyra X6. If you are evaluating it today, you are evaluating it for a few months of use or as a reference point for its successor.

What X5 was built for

WRITER announced Palmyra X5 on 28 April 2025, available on its own platform and on Amazon Bedrock from launch. It reached OpenRouter's catalogue much later, which is why some listings date it to January 2026. The announcement positions it for enterprise agents that work across large volumes of company data and many tools.

The examples WRITER gives are telling: drafting RFP responses, generating fund reports, classifying support tickets, analysing long contracts, 10-K filings and medical records, and summarising large sets of survey responses or papers. Every one of them is read-heavy. The model ingests a lot and produces a comparatively short result.

On architecture, the announcement describes a hybrid attention design that blends softmax and linear attention, combined with a mixture-of-experts structure. It does not give a parameter count, and it is offered as a hosted model through WRITER's API and Bedrock rather than as downloadable weights.

The price shape: cheap to read, expensive to write

WRITER lists Palmyra X5 at $0.60 per million input tokens and $6.00 per million output through its API. OpenRouter listed the same rates at the time of writing. WRITER's own agent product bills it differently, at $5.00 in and $12.00 out, so check which surface you are pricing.

The ten-to-one ratio decides the economics. A 500K-token contract with a 1,000-token answer costs about $0.30 in input and well under a cent in output. Flip the ratio and the model is expensive for what it does. Input vs output token pricing covers how to tell which side of the bill your workload lands on.

WRITER also claims speed at that scale: about 22 seconds to process a full 1M-token prompt, and multi-turn function calls in around 300 milliseconds. Those are vendor figures; measure time to first token on your own long prompts before building a user-facing flow around them.

8,192 output tokens is the constraint

The output cap is the limit you will hit first. 8,192 tokens is a few thousand words of prose, or a modest code file. It rules out generating long reports in one call, producing large patches, or letting a reasoning-style model think at length before answering.

The workarounds are standard but cost turns: generate section by section, or have X5 produce a structured outline and findings that a cheaper model expands. If your task needs long output by nature, X5 is the wrong model regardless of its window.

One more practical detail. WRITER describes X5 as built for tool-using agents, but OpenRouter's listing for writer/palmyra-x5 accepts only max_tokens, stop, temperature, top_k and top_p, with no tools parameter. If you need function calling, use WRITER's API or Bedrock directly and confirm it works there.

Vendor benchmark claims, read carefully

WRITER's announcement reports these results:

• OpenAI MRCR, 8-needle: 19.1%, against 20.25% for GPT-4.1 and 17.63% for GPT-4o.
• BigCodeBench (Full, Instruct): 48.7.
• BBH: 70.99%, GPQA: 47.20%, MMLU-Pro: 65.02%, MATH Hard: 71.57%.

The MRCR figure is the relevant one, because it tests retrieval of several items from a very long context, which is X5's job. Note what it says: even the leading score in that comparison is around one in five. Retrieval at this depth was hard for every model in that comparison, so plan on verification rather than trust. Long context degradation explains why quality falls well before the advertised limit.

The general reasoning scores are modest by 2026 standards, and the comparisons are against GPT-4.1 and GPT-4o. They describe a model built for document throughput, not for frontier reasoning.

The deprecation date, and X6

WRITER's documentation lists Palmyra X6 as the migration target and current default in its agent product: 1M context, the same 8,192-token output cap, at $2.00 in and $8.00 out per million tokens. Palmyra X4, with a 128K window, is scheduled for deprecation on the same December date.

X6 costs more than three times as much on input. For read-heavy workloads, where input dominates, that is the number that will move your bill. Budget for it now rather than in December. Model deprecation and migration covers how to plan a forced move without a rushed cut-over.

When to still pick X5

• Pick X5 for the next two months if you already run it on long-document extraction or summarisation and the outputs are short. Use the time to replay your traffic on X6 and alternatives.
• Do not start a new integration on X5 unless it is a short-lived project that will finish before 14 December.
• Compare widely before migrating. Several 2026 models offer 1M windows with far larger output limits; best model for long context tasks covers how to match one to the shape of your work.