← Library · Model comparisons Model comparison · cheat sheet

AI model cheat sheet, November 2026 editionWhat the 9 models we use at work are good at, and what they cost

This page summarizes the 9 models Dorado uses for research, writing, and development, across 3 vendors: Anthropic, OpenAI, and Google. For each model, we separate what the vendor says it does best from what we saw when we gave several models the same task. Prices and third-party benchmark figures were checked on 21 Nov 2026.

No single model is best at everything. For important research where thorough coverage matters, we use more than one model.

21 Nov 2026

Our take
  • No single model is best at everything.
  • For important research where thorough coverage matters, we use more than one model (20 Nov 2026 comparison).
  • When we hit a wall, we try another model to see whether a different angle comes out (18 Nov 2026 comparison).
  • Every model still needs a source check. In the 18 Nov test, 3 of 30 cited passages had numbers or headings that did not match the source, and in the 20 Nov test, 6 of 19 did.
Cheat sheet

What the 9 models do best, and what our tests showed

Vendor is the vendor's description. Our test is what we saw when we sent the same Japanese research request to several models. We compared only research tasks that use web search, not coding. Every model ran at the "high" reasoning level.

TestedTested with an earlier version (GPT-6 Astra, GPT-6.1 Sol)Not testedInput priceOutput price (right end of bar is $40)Vals rank strip: #1 on the left, #22 on the right

ModelVendorVendor descriptionOur testRuns tested
11/1211/1411/1811/20
Our testWhat our tests showedAPI priceInput / OutputDefault
reasoning
Vals rank
Claude Opus 5.5AnthropicVendorLong-running agentic coding and knowledge work.
11/1211/1411/1811/20
Our test14 Nov: Strong at deep research. Too thorough for broad discovery.
$4 / $20
medium
#5
Claude Sonnet 5.5AnthropicVendorBest mix of speed and intelligence. Faster, cheaper companion to Opus 5.5.
11/1211/1411/1811/20
Our test14 Nov: Broad, low-cost research. 28 min, 48 facts, estimated cost $2.35.
$2 / $10
high
#3
Claude Fable 5.2AnthropicVendorHard reasoning and long-running agentic work.
11/1211/1411/1811/20
Our test18 and 20 Nov: Most facts in its notes, and also the most time and tokens.
$8 / $40
high
#1
Claude Haiku 5.5AnthropicVendorHigh-volume, fast tasks such as classification, extraction, and routing.
11/1211/1411/1811/20
Our testNot tested.
$0.10 / $0.50Up to 100K tokens
medium
#22
GPT-6.2 AstraOpenAIVendorOpenAI's most capable model. For difficult projects from start to finish.
11/12 (earlier version)11/14 (earlier version)11/1811/20
Our test18 and 20 Nov: Fastest in both runs (7 min 30 s, 6 min 40 s).
$8 / $40
mediumOpenAI recommended
#8
GPT-6.2 SolOpenAIVendorThe everyday workhorse. From coding to high-volume agents.
11/1211/14 (earlier version)11/1811/20
Our test18 Nov: 8 min 24 s, 2.67M tokens. All 6 cited passages were backed.
$2 / $10
mediumOpenAI recommended
#10
GPT-6.1 TerraOpenAIVendorFast, efficient model for visual reasoning, computer use, and natural writing.
11/1211/1411/1811/20
Our testNot tested.
$1 / $5
medium
#17
Gemini 4 NeonGoogleVendorComplex coding, analytical reasoning, and production enterprise work.
11/1211/1411/1811/20
Our test20 Nov: Most readable writing, and hit the key points.
$3 / $15
high
#2
Gemini 4 ArgonGoogleVendorPractical coding, enterprise knowledge work, and cyber defense.
11/1211/1411/1811/20
Our testNot used.
$2 / $10Until 30 Nov. $4 / $20 from 1 Dec
Unknown
#6

API prices are standard rates in US dollars per 1M tokens. Vals ranks are as shown on 19 Nov 2026. All 9 models take a context window of about 1M tokens. Figures from our tests come from 1 run per model in each comparison.

By task

Which model to try first, by task

We regrouped the comparison results above by task. For models we have not tested, we follow the vendor's description.

  • Get an overview fastGPT-6.2 Astra. It finished fastest in both the 18 Nov and 20 Nov comparisons.
  • Gather the most factsClaude Fable 5.2. It put the most facts in its notes, and also used the most time and tokens.
  • Dig deep into a fixed scopeClaude Opus 5.5. To keep costs lower, try GPT-6.2 Sol (the 14 Nov test used the earlier 6.1 Sol).
  • Turn findings into a readable reportGemini 4 Neon. Its writing was the easiest to read in the 20 Nov comparison. One option is to research with several models and then edit the report with Neon.
  • High-volume classification or extractionClaude Haiku 5.5 or GPT-6.1 Terra. Based on vendor descriptions; we have not tested either.
  • Important researchSend the same request to more than one model. When you hit a wall, try another model.
By model

Descriptions and test results, model by model

Here are the full notes behind each row of the cheat sheet above.

Claude Opus 5.5Anthropic · 2026-09-22 $4 / $20 · medium

VendorBuilt for long-running agentic coding and knowledge work. Matches Fable 5.1 on many tasks.

Our test14 Nov: Strong at deep research. It logged every citation location and pulled profit margins from financial filings. Too thorough for broad discovery.

18 Nov: 9 min 13 s, 4.56M tokens, 13 facts in the note.

Claude Sonnet 5.5Anthropic · 2026-09-28 $2 / $10 · high

VendorBest mix of speed and intelligence. A faster, cheaper companion to Opus 5.5. Anthropic itself states that Opus 5.5 is stronger on complex, open-ended work.

Our test14 Nov: Broad, low-cost research. 28 min, 48 facts, estimated cost $2.35.

In our Claude runs, it also served as the child task handling web searches and page reads.

Claude Fable 5.2Anthropic · 2026-11-09 $8 / $40 · high

VendorBuilt for hard reasoning and long-running agentic work. Even so, Anthropic advises starting with Opus 5.5 for most tasks.

Our test12 Nov: Detailed coverage of 31 trade shows, down to visitor breakdowns and exhibitor names. Used 5.8x as many tokens as Astra.

14 Nov: Strong at broad, fast research. 16 min, 98 facts.

18 and 20 Nov: Put the most facts in its notes, and also used the most time and tokens (18 min 20 s and 7.17M; 17 min 45 s and 7.64M).

Claude Haiku 5.5Anthropic · 2026-10-07 $0.10 / $0.50 (up to 100K tokens) · medium

VendorBuilt for high-volume, speed-sensitive tasks such as classification, extraction, and routing. Anthropic recommends Sonnet 5.5 or Opus 5.5 for complex agentic coding.

Our testNot tested.

GPT-6.2 AstraOpenAI · 2026-11-17 (reported) $8 / $40 · medium (OpenAI recommended)

VendorOpenAI's most capable model. Intended for difficult projects from start to finish.

Our test18 Nov: Fastest at 7 min 30 s and 2.58M tokens. It went looking for what client-side and vendor-side contacts actually say and do with each other, which the other 3 models did not look for.

20 Nov: Fastest at 6 min 40 s, and all 5 cited passages were backed.

The earlier GPT-6 Astra cast a wide net of 54 trade-show candidates on 12 Nov, and got a quick overview on 14 Nov.

GPT-6.2 SolOpenAI · 2026-11-17 (reported) $2 / $10 · medium (OpenAI recommended)

VendorThe everyday workhorse. From coding to high-volume agents.

Our test18 Nov: 8 min 24 s, 2.67M tokens. All 6 cited passages were backed.

The earlier GPT-6.1 Sol ran deep research at low cost on 14 Nov: 48 min, estimated cost $0.83.

GPT-6.1 TerraOpenAI · 2026-10-26 (reported) $1 / $5 · medium

VendorA fast, efficient model for visual reasoning, computer use, and natural writing. OpenAI itself states that GPT-6.1 Sol and GPT-6 Astra are better for the most complex coding.

Our testNot tested.

Gemini 4 NeonGoogle · 2026-11-18 $3 / $15 · high

VendorBuilt for complex coding, analytical reasoning, and production enterprise work. In Google's lineup, it sits between Carbon (the hardest long-running tasks) and Hydrogen (fast everyday tasks).

Our test20 Nov: Produced the most readable writing and hit the key points. 13 min 35 s, 4.60M tokens. 4 of 7 cited passages were fully backed (the same share as Fable).

Gemini 4 ArgonGoogle · 2026-09-30 $2 / $10 (until 30 Nov; $4 / $20 from 1 Dec) · unknown

VendorBuilt for practical coding, enterprise knowledge work, and cyber defense. Responds faster than Carbon and uses fewer tokens per turn. Maximum output is 1M tokens, up from 64K in earlier Gemini models.

Our testNot used.

Reasoning levels

Higher reasoning levels take more time and money, and do not always bring matching gains

All 3 vendors let you choose in steps how much a model thinks before answering. Tokens spent thinking are billed at the output rate.

VendorLevelsVendor guidance
Anthropiclow, medium, high, xhigh, max

low saves a large share of tokens in exchange for a small drop in capability. high uses as many tokens as the task requires. xhigh is for agentic work and coding that run past 30 minutes.

max is slower and more expensive than xhigh, and the extra gain is often small. Fable 5.2 already delivers most of its capability gain at high.

OpenAIlow, medium, high, xhigh, maxTerra also offers none

Start new workloads at medium, which balances speed, quality, and cost. Use high when quality matters more than speed.

Use xhigh only when testing shows a gain worth the extra time and cost.

GoogleLow, Medium, High, Extra HighNeon. Not documented for Argon

Capability rises steadily with each step. High is the default. Extra High adds further rounds of thinking for the hardest tasks.

Third-party benchmarks

Scores at the top of Vals are close, while cost runs from $0.41 to $26.54 per task

Figures from 3 benchmarks. Each reflects a point in time, and rankings shift month by month. Artificial Analysis cost is in US dollars per task. Vals cost is per test.

Vals Index score

Right end of bar is 100%. Ordered by Vals rank

Fable 5.2
71.31% (#1)
Gemini 4 Neon
70.87% (#2)
Sonnet 5.5
70.85% (#3)
Opus 5.5
70.14% (#5)
Gemini 4 Argon
69.39% (#6)
GPT-6.2 Astra
67.66% (#8)
GPT-6.2 Sol
66.06% (#10)
GPT-6.1 Terra
60.54% (#17)
Haiku 5.5
58.73% (#22)

Artificial Analysis cost (per task)

Right end of bar is $26.54. Ordered by Vals rank

Fable 5.2
$26.54
Gemini 4 Neon
$6.44
Sonnet 5.5
$10.02
Opus 5.5
$10.04
Gemini 4 Argon
$9.95
GPT-6.2 Astra
$4.00
GPT-6.2 Sol
$1.01
GPT-6.1 Terra
$0.96
Haiku 5.5
$0.41
ModelArtificial Analysis
Score (reasoning) · Cost
Vals Index
Rank · Score · Cost
LMArena (Text)
Rank
Claude Fable 5.265 (max) · $26.54#1 · 71.31% · $17.94#7
GPT-6.2 Astra62 (max) · $4.00#8 · 67.66% · $10.68Outside top 15
Gemini 4 Neon61 (xhigh) · $6.44#2 · 70.87% · $7.40#3
Claude Opus 5.559 (max) · $10.04#5 · 70.14% · $24.77#5
GPT-6.2 Sol57 (max) · $1.01#10 · 66.06% · $3.18Outside top 15
Claude Sonnet 5.555 (max) · $10.02#3 · 70.85% · $16.18Outside top 15
GPT-6.1 Terra51 (max) · $0.96#17 · 60.54% · $2.94Outside top 15
Gemini 4 Argon50 (high) · $9.95#6 · 69.39% · $13.37#2
Claude Haiku 5.536 (max) · $0.41#22 · 58.73% · $3.17Outside top 15

Vals as shown on 19 Nov 2026; LMArena as shown on 20 Nov 2026. The Artificial Analysis page carries no date. All checked on 21 Nov 2026.

In the Vals breakdown, every model scored low on the legal benchmark (HLAB), with Fable 5.2 at 5.0%. On the tax benchmark (Tax Agent Bench), Astra scored 21.8% and Sol 24.8%, well below Neon's 54.9%. On Vals's coding benchmark (Terminal-Bench 4), Argon scored lowest among the 9 models at 57.6%.

Cautions

Things to know before using them

  • In the Gemini app's model picker, Argon was reportedly replaced by Neon (9to5Google, 18 Nov 2026). Argon's API price doubles on 1 Dec.
  • Users have reportedly complained that GPT-6.2 Astra stops partway through long tasks to ask odd questions or repeats the same step. An OpenAI representative said changes added in the last 3 or 4 days before launch were not retested (Wccftech).
  • Anthropic itself states that Claude Opus 5.5 and Fable 5.2 sometimes suspect they are being tested, which makes their behavior in real use harder to predict.
  • Most benchmark scores published by vendors are measured in-house. The table above includes only 3 third-party benchmarks.
  • Claude Haiku 5.5 counts the same text as about 30% more tokens than Haiku 4.5.
Limits of this cheat sheet

4 comparisons, all research on Japan

  • Our results come from 4 comparisons, with 1 run per model in each. Every topic was research on Japan using web search. We did not compare coding, images, or long-document writing.
  • The 12 Nov and 14 Nov comparisons used the earlier versions, GPT-6 Astra and GPT-6.1 Sol.
  • Model names, prices, and rankings change month by month. This sheet reflects values as of 21 Nov 2026.
Comparison articles

Comparison articles behind this cheat sheet

DateTopicModels compared
2026-11-20Communication styles that leave a good impression in Japanese businessGemini 4 Neon, Claude Fable 5.2, GPT-6.2 Astra
2026-11-18Workplace and business-partner relationships in JapanGPT-6.2 Astra, GPT-6.2 Sol, Claude Opus 5.5, Claude Fable 5.2
2026-11-14Japan's growth industries by CAGRClaude Fable 5.2, GPT-6 Astra, Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol
2026-11-12Major trade shows in JapanClaude Fable 5.2, GPT-6 Astra
About this cheat sheet

Check date, scope, and sources

Checked
21 Nov 2026
Scope
9 models: Claude Opus 5.5, Sonnet 5.5, Fable 5.2, Haiku 5.5, GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, Gemini 4 Neon, Gemini 4 Argon
Selection
The 8 models other than Gemini 4 Argon are the ones we used most over the past 30 days. We added Argon as Google's top model released on 30 Sep
Sources
Official vendor pages, 3 third-party benchmarks (Artificial Analysis, Vals, LMArena), and 4 of our own comparison articles
Pricing
Standard API rates, in US dollars per 1M tokens
Editing
Vendor descriptions are summarized from the originals. You can open the originals from the sources below
Sources

Sources (checked 21 Nov 2026)

The research plan
$50 first month, then $200 a month

Want the sources checked for your product?

These answers are unchecked. In the research plan, we open the sources, check the dates, and correct what is wrong before the brief reaches you.