Hardcore Reviews
Hardcore Reviews

Flagships Plan, But Who Executes? Four Lightweight Ledgers

"Let the flagship plan, let a lightweight tier execute" is the shared agent-engineering story of 2026 - but the ledgers are not uniform. As of November 10, 2026 four lightweight execution tiers are live at once: Meta Muse Ember (muse-ember-1.0, $0.10 in / $0.50 out per million tokens, $0.30/$1.50 above 200K input, five modalities, 1,048,576 context), Anthropic Haiku 5.5 (verbatim positioning "Fastest, most cost-efficient model", segmented $0.10-$0.50 / $0.50-$2.50 by prompt under/over 100K, modality and context not stated on its pricing page), Google Gemini 3.5 Flash-Lite ("The fastest, lowest-cost model in the 3.5 family", $0.30/$2.50, 1M context, five modalities; the Gemini 3.8 Flash introductory $0.75/$3.75 runs only until December 31, 2026) and OpenAI GPT-6 Luna (gpt-6-luna, "Our most efficient model for focused, high-volume tasks", $0.10/$0.50, 1.05M context, whole-request surcharge above 272K; GPT-5.4 mini/nano are deprecated). The four ledgers are lined up with verbatim positioning, registered modalities and context, and pricing with all qualifiers attached - zero benchmarks, zero ranking, zero recommendation; Meta's own cross-vendor scoreboard is not reproduced and the Muse Code subscription price stays out of the API comparison.

Published November 11, 20269 min read
<!-- lightweight-agent-models-comparison-review | review | Flagships Plan, But Who Executes? Four Lightweight Ledgers -->

There is one story every vendor tells in 2026: let the flagship plan, let a lightweight tier execute. The main agent decomposes the task, routes tools, and recovers from errors; underneath it sits a row of small models running coding subtasks, tool loops, and document extraction. The narrative is uniform. The ledgers are not. As of November 10, 2026, four lightweight execution tiers are live at the same time - Meta Muse Ember, Anthropic Haiku 5.5, Google Gemini 3.5 Flash-Lite, and OpenAI GPT-6 Luna - and the same job of "doing the work" is priced from $0.10 to $0.30 per million input tokens and $0.50 to $2.50 per million output, with modality coverage ranging from blanks marked "not listed" to five modalities open. This article lines the four ledgers up side by side: verbatim positioning, modality and context registration, and pricing with every qualifier attached. No benchmarks, no rankings, no recommendations.

On scope and sourcing first. This piece is a selection-profile comparison of the lightweight execution tier, not a full price table: the complete per-token breakdown, cache hit rates, and aggregator fees live in our LLM API pricing comparison, a deliberately different angle. On verification: Anthropic's claude.com/pricing page was fetched directly; the official pages of Meta, Google, and OpenAI resisted direct fetches, so figures below come from search excerpts restricted to official domains, cross-checked against official page wording first.

One label each: how the four describe "the model that does the work"

Meta's Muse Ember model page on dev.meta.ai reads: "Muse Ember is our smallest multimodal model built for high-volume agentic workflow execution, from coding subtasks and tool calls to document summarization and extraction." The model ID is muse-ember-1.0, officially marked available in Standard mode only. Meta's model-selection documentation sketches the use cases: sub-agent tasks handed out by a main agent, including coding subtasks; tool calls, tight tool loops, and latency-sensitive turns; classification, routing, and other high-volume requests; the same task run over thousands of images, files, or records. The capability language stays in workflow territory: coding subtasks, tool calls, routing, document processing.

Anthropic's claude.com/pricing gives Haiku 5.5 a seven-word label: "Fastest, most cost-efficient model." No further ornament, but the same page supplies the context: it is the only Haiku with a positioning line there, the previous Haiku 4.5 ($1 input, $5 output) is marked Legacy, and the flagship ladder reads Sonnet 5.5 at $2/$10, Opus 5.5 at $4/$20, and Fable 5.2 at $8.00/$40.00 Standard with Fast at twice that. The lightweight-versus-flagship staircase sits on a single vendor page.

Google's Gemini 3.5 Flash-Lite model page states verbatim: "The fastest, lowest-cost model in the 3.5 family. Outperforms prior Flash-Lite generations for high-throughput execution." Two timeline details deserve side-by-side registration: the model ID gemini-3.5-flash-lite carries a Stable status with a "Latest update July 2026" note, while the 3.6 Flash and 3.5 Flash-Lite pages are flagged as previous-generation releases; no newer Flash-Lite generation had appeared as of the verification date. The page holds both facts without resolving them.

OpenAI's GPT-6 Luna model page reads: "Our most efficient model for focused, high-volume tasks." The model index pairs it with a sibling: "Choose GPT-6.1 Sol to balance intelligence and cost, or GPT-6 Luna for cost-sensitive, high-volume workloads." The model ID is gpt-6-luna.

Modalities and context windows: register what the pages say

What each tier can ingest differs more than the list prices do.

ModelOfficial positioning (verbatim excerpt)Input modalitiesContext window
Meta Muse Embersmallest multimodal model for high-volume agentic workflow executionText, image, video, audio, PDF1,048,576 tokens
Anthropic Haiku 5.5Fastest, most cost-efficient modelNot listed on the verified pageNot listed on the verified page
Google Gemini 3.5 Flash-LiteThe fastest, lowest-cost model in the 3.5 familyText, image, video, audio, PDF1,048,576 input, 65,536 output
OpenAI GPT-6 LunaOur most efficient model for focused, high-volume tasksText, image (text output)1,050,000 (about 1.05M)

Muse Ember takes all five input types - "Text, image, video, audio, and PDF" - pitched as bringing multimodal understanding to high-volume agent workflows; the 1,048,576-token context window comes from the model-selection table. Gemini 3.5 Flash-Lite matches the five-modality input with text-only output; the window is 1,048,576 input tokens with a 65,536-token output cap. GPT-6 Luna accepts text and image only; audio and video are marked "Not supported" on the model page, the context window is 1,050,000 tokens (about 1.05M), output tops out at 128,000 tokens, and the knowledge cutoff is May 18, 2026. Haiku 5.5 is the only entry with blanks in both columns: the directly fetched pricing page does not list its input modalities or context limit, and the model documentation page was region-blocked during verification. The article registers "not listed on the official page" rather than filling the gap cross-vendor.

Capability granularity differs as well. The Gemini 3.5 Flash-Lite page itemizes support: Batch API, caching, code execution, function calling, structured outputs, and thinking; audio generation, image generation, and the Live API are marked not supported. GPT-6 Luna supports web search, file search, the code interpreter, computer use, and MCP, with fine-tuning not supported. Muse Ember's capability description leans on workflow vocabulary rather than a feature matrix. Haiku 5.5's capability list falls outside the pages verified here and stays unwritten.

The pricing ledger: numbers travel with their qualifiers

The list prices look converged - $0.10 to $0.30 input, $0.50 to $2.50 output - but three of the four bury staircase or cache qualifiers under the headline numbers. Copying only the floor price miscalculates the bill. Take them in turn.

Muse Ember splits at prompt length: at or under 200K input tokens, cached input $0.004, input $0.10, output $0.50; past 200K, cached input $0.012, input $0.30, output $1.50. The critical clause is how the upgrade applies: a request with more than 200K input tokens is "billed at the long-context rates in full, not only for the tokens past 200K" - the real unit price of a long prompt jumps for the entire request, not just the overflow. Contributor mode is not available for Muse Ember on the Model API; dev.meta.ai also lists Muse Code at $5.00/month, a subscription price for a different product that stays out of this API comparison.

Haiku 5.5 segments by prompt length: input $0.10 up to 100K tokens and $0.50 past that; output $0.50 up to 100K and $2.50 past it. Cache pricing segments the same way: read $0.01 and write $0.125 within 100K, read $0.05 and write $0.625 beyond. Unlike Muse Ember's whole-request jump, Anthropic presents both columns side by side on one page - a five-fold output-price spread inside a single lightweight tier.

Gemini 3.5 Flash-Lite is the only flat single-tier quote of the four: $0.30 per million input tokens and $2.50 per million output tokens, with a default thinking level of "minimal." The pages verified here list no cache unit price or long-context staircase; those cells stay blank rather than invented. One timeline entry belongs to the sibling Gemini 3.8 Flash: its introductory price of $0.75/$3.75 runs only through December 31, 2026, moving to a standard $1.50/$7.50 from January 1, 2027 - a limited-time price printed on the official page, recorded without extrapolation.

GPT-6 Luna's sticker prices match Muse Ember's floor: input $0.10, cached input $0.01, cache writes $0.125, output $0.50. The long-prompt rule is finer: prompts above 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request. The page also lists Batch/Flex at 50 percent off and Fast mode at 2x. All four tiers, with qualifiers in the same cells:

ModelInput priceOutput priceStaircase and cache qualifiers (same row)
Muse Ember$0.10 ($0.30 above the 200K tier)$0.50 ($1.50 above the 200K tier)Cached input $0.004 and $0.012; above 200K billed at long-context rates in full
Haiku 5.5$0.10 ($0.50 for prompts past 100K)$0.50 ($2.50 for prompts past 100K)Cache read $0.01, write $0.125 within 100K; read $0.05, write $0.625 past it
Gemini 3.5 Flash-Lite$0.30$2.50Single-tier quote; no cache price or staircase on the pages verified
GPT-6 Luna$0.10 (2x for the whole request past 272K)$0.50 (1.5x for the whole request past 272K)Cached input $0.01, cache writes $0.125; Batch/Flex 50% off, Fast mode 2x

Gaps and historical footnotes: the unwritten parts matter too

The two blanks under Haiku 5.5 deserve careful handling. "Not listed on the official pricing page" is a statement about the page, not about the model; the documentation page was region-blocked during verification, so the registration stays at "not listed" rather than hardening into "officially unsupported." A labeled information-access gap is cleaner than paraphrasing the vendor into a claim it never made.

Historical footnotes matter as well. On the OpenAI side, GPT-5.4 mini ($0.75/$4.50, 400K context) and GPT-5.4 nano ($0.20/$1.25) are both marked Deprecated on the model pages, and the earlier GPT-5 nano is deprecated too, its page steering most new speed- and cost-sensitive workloads to GPT-6 Luna. GPT-5.6 Luna, at $0.20/$1.20, "roughly corresponds to the nano model tier used in earlier GPT-5 families" in OpenAI's own description. The lineage: Luna inherited the nano seat; the mini and nano names are history. On the Google side, the 3.8 Flash introductory price and the previous-generation flags coexist, so a selection needs three fields checked together: which generation, which price, until when.

One section is deliberately missing: benchmarks. Meta's dev.meta.ai pages carry a cross-vendor score table that includes Haiku 5.5 and GPT-6 Luna. This column's rule is zero benchmarks, zero rankings, and no cross-table comparison, so the article registers the table's existence and copies none of its numbers; readers who want scores can visit Meta's model-selection documentation directly.

Three questions before you commit

The table answers "what is it"; selection asks "does it fit." Three question groups work with official wording alone. First, modality fit: bulk document extraction and multimodal triage need the five-modality inputs of Muse Ember or Gemini 3.5 Flash-Lite; GPT-6 Luna's text-plus-image pair covers common screenshot and table workloads but admits no audio or video; Haiku 5.5's modality coverage is unlisted, so confirm against official documentation before committing. Second, context length against staircase pricing: three of the four change the price at a length threshold - Muse Ember at 200K for the whole request, GPT-6 Luna above 272K at 2x input and 1.5x output, Haiku 5.5 segmenting at 100K. Which side your typical prompt lands on decides whether the effective input price reads $0.10 or $0.30, the output $0.50 or $2.50. Third, cache dependence: every cache price sits well below the sticker input price ($0.004 to $0.125), and in high-concurrency agent workflows cache hit rates amplify into bill-level differences; the full arithmetic lives in our pricing comparison.

Run the three groups and the ledger moves from "side by side" to "on demand." Within one tier called lightweight, Meta bets on multimodal throughput, Google on low latency at low cost, OpenAI on a minimal sticker for focused tasks, and Anthropic hands the choice back with a seven-word label and a five-fold segmented spread. Which tier fits your workflow depends on prompt-length distribution, modality mix, and cache hit rate - not on any leaderboard. That is the point of writing without scores or recommendations.

FAQ

Q1: How large are the context windows across the four tiers? A1: Meta Muse Ember lists 1,048,576 tokens (about 1M); Google Gemini 3.5 Flash-Lite lists 1,048,576 input and 65,536 output; OpenAI GPT-6 Luna lists 1,050,000 (about 1.05M) with 128,000 maximum output. Haiku 5.5's context limit is not listed on the official pricing page verified here, not filled by guesswork.

Q2: Why are Haiku 5.5's modality and context fields blank? A2: The directly fetched claude.com/pricing page does not list input modalities or a context limit for Haiku 5.5, and the model documentation page was region-blocked during verification. Following the rule that "not listed" is not "officially unsupported," the article registers the gap and defers to official documentation.

Q3: Which long-prompt surcharges apply? A3: Muse Ember bills requests above 200K input tokens at long-context rates in full (input $0.30, output $1.50, cached input $0.012); GPT-6 Luna prices prompts above 272K input tokens at 2x input and cache rates and 1.5x output for the full request; Haiku 5.5 segments at 100K tokens ($0.50 versus $2.50 output); Gemini 3.5 Flash-Lite shows no staircase on the pages verified. The three surcharge rules apply to the whole request, not only the overflow.

Q4: Meta's official page has a score table - why is it not reproduced here? A4: dev.meta.ai does carry a cross-vendor score table that includes Haiku 5.5 and GPT-6 Luna. This column runs on zero benchmarks, zero rankings, and no cross-table comparison, so it registers the table's existence and copies no numbers; readers who want scores can open the official page directly.

Q5: Are GPT-5.4 nano and mini still options? A5: Both are marked Deprecated on OpenAI's model pages and register as history rather than choices: mini at $0.75/$4.50 with 400K context, nano at $0.20/$1.25. The earlier GPT-5 nano is deprecated as well, its page steering new cost-sensitive workloads to GPT-6 Luna; GPT-5.6 Luna ($0.20/$1.20) roughly corresponds to the earlier nano tier in OpenAI's own description.


Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-11

FAQ

How large are the context windows across the four tiers?
Meta Muse Ember lists 1,048,576 tokens (about 1M); Google Gemini 3.5 Flash-Lite lists 1,048,576 input and 65,536 output; OpenAI GPT-6 Luna lists 1,050,000 (about 1.05M) with 128,000 maximum output. Haiku 5.5's context limit is not listed on the official pricing page verified here, not filled by guesswork.
Why are Haiku 5.5's modality and context fields blank?
The directly fetched claude.com/pricing page does not list input modalities or a context limit for Haiku 5.5, and the model documentation page was region-blocked during verification. Following the rule that "not listed" is not "officially unsupported," the article registers the gap and defers to official documentation.
Which long-prompt surcharges apply?
Muse Ember bills requests above 200K input tokens at long-context rates in full (input $0.30, output $1.50, cached input $0.012); GPT-6 Luna prices prompts above 272K input tokens at 2x input and cache rates and 1.5x output for the full request; Haiku 5.5 segments at 100K tokens ($0.50 versus $2.50 output); Gemini 3.5 Flash-Lite shows no staircase on the pages verified. The three surcharge rules apply to the whole request, not only the overflow.
Meta's official page has a score table - why is it not reproduced here?
dev.meta.ai does carry a cross-vendor score table that includes Haiku 5.5 and GPT-6 Luna. This column runs on zero benchmarks, zero rankings, and no cross-table comparison, so it registers the table's existence and copies no numbers; readers who want scores can open the official page directly.
Are GPT-5.4 nano and mini still options?
Both are marked Deprecated on OpenAI's model pages and register as history rather than choices: mini at $0.75/$4.50 with 400K context, nano at $0.20/$1.25. The earlier GPT-5 nano is deprecated as well, its page steering new cost-sensitive workloads to GPT-6 Luna; GPT-5.6 Luna ($0.20/$1.20) roughly corresponds to the earlier nano tier in OpenAI's own description. ---

Related

Hardcore Reviews

Small Model Price War: Haiku 5.5 Ties Luna Only Below 100K

A small-model price-war comparison (October 8, 2026 price snapshots; five columns, four rows): Claude Haiku 5.5, GPT-6 Luna, Sonnet 5.5, and Qwen3.8-27B as the open self-hosting row. Pricing: Haiku 5.5 at $0.10/$0.50 per million (under-100K prompts), Luna $0.10/$0.50 (up to 272K), Sonnet 5.5 at $2/$10 list with cache reads halved to $0.10, and Qwen3.8-27B free weights (Apache-2.0) on your own GPU. Core finding: Haiku 5.5 matches Luna only in the short-prompt tier - above 100K tokens its $0.50 input faces Luna's $0.20, so long-document batch bills invert the picture. Benchmarks cite each model's own official numbers, labeled and never mixed across suites: Haiku 5.5 Terminal-Bench 4.0 at 39.2% and OSWorld 72.4%; no fresh Luna numbers shipped with the release; Sonnet 5.5 at 70.6% on TB4; Qwen3.8-27B has no cross-suite table. Effort settings compared (Haiku 5.5 is the first Haiku-class with the dial; Luna runs none-to-max), contexts (1M for both hosted models; Qwen 262K extensible to 1M). Verdicts by use case - high-volume subagents, batch classification, main-line coding, private deployment - with the site's standing no-stress-test-no-arbitration stance. Four same-topic deep dives linked inline.

Oct 8, 202610 min read
Hardcore Reviews

TrnRT, Kindi, Baseten: One License Forbids Managed Service

Three inference engine repositories, TrnRT, Kindi and Baseten, on one comparison table, with license terms and hardware bets as the axis rather than a buying recommendation or benchmark ranking. Apache-2.0, Elastic License 2.0 and AGPL-3.0 coexist; one LICENSE says verbatim that the software may not be provided to third parties as a hosted or managed service, and GitHub reports NOASSERTION. Hardware claims from NKI megakernels on Trainium2 to GB300, B300, Quadro K420 and RTX 3080 are labelled as README self-reports. Stars are cumulative totals, with no growth score or single-day delta.

Nov 16, 20269 min read