Radar
← All Radar guides

Radar

Ship checklist

The hype is easy. Shipping is hard. These are the practical checks and gotchas that determine if a model will actually survive in production.

10 models

GPT-6 Astra

OpenAI ·

APIMultimodalReasoning
Context Window
1M
License
Proprietary (API)
Pricing
$10 input / $50 output per 1M
Fine-tune
noPrompt and tool steering only. Fine-tuning platform is winding down.

Best for

  • › Hardest coding and research
  • › Long-horizon agents
  • › When 5.6 Sol is not enough

Strengths

  • › Current OpenAI flagship
  • › 1M context / 128K max output
  • › Strong agentic and STEM work

What usually breaks

  • ! Rollout is staged; API access may still be gated on some accounts
  • ! Cyber-sensitive capabilities are trusted-access only
  • ! Fast mode is 2x Standard price
Docs / source →

Claude 5 Fable

Anthropic ·

APIMultimodalReasoning
Context Window
200K+
License
Proprietary (API)
Pricing
$10 input / $50 output per 1M
Fine-tune
noSteer via system prompts and tools.

Best for

  • › High-stakes accuracy
  • › Complex coding (SWE-bench champ)
  • › Careful enterprise tasks

Strengths

  • › Best overall correctness model
  • › Leads SWE-bench Verified (95.0%)
  • › Careful reasoning

What usually breaks

  • ! Safety refusals remain strict
  • ! Recent 18-day government suspension impacted availability
Docs / source →

Kimi K3

Moonshot ·

Open weightsReasoning
Context Window
Very long
License
Modified MIT
Pricing
Free weights; GPU cost for serving
Fine-tune
yesLargest open-source model (2.8T); tuning requires massive compute.

Best for

  • › Enterprise private deployments
  • › Frontier-level reasoning on-prem

Strengths

  • › Largest open model (2.8T)
  • › Rivals GPT-5.6 and Claude Fable 5
  • › Leads open tool use (BrowseComp)

What usually breaks

  • ! Massive parameter count requires extreme infra to serve locally
  • ! Multimodal support unclear
Docs / source →

Inkling

Thinking Machines ·

Open weightsMultimodalReasoning
Context Window
1M
License
Open-weights
Pricing
Free weights; Tinker API options available
Fine-tune
yesMoE architecture requires specific fine-tuning approaches.

Best for

  • › Custom multimodal models
  • › Efficient long-context reasoning

Strengths

  • › 975B MoE (41B active)
  • › Controllable thinking effort
  • › Native multimodal

What usually breaks

  • ! Brand new ecosystem (Mira Murati's lab)
  • ! MoE routing can complicate deterministic outputs
Docs / source →

GPT-5.6 Sol

OpenAI ·

APIMultimodalReasoning
Context Window
Large (varies by SKU)
License
Proprietary (API)
Pricing
$4 input / $20 output per 1M (promo through 2026-11-21)
Fine-tune
noPrompt and tool steering only for now.

Best for

  • › Multi-agent parallel tasks
  • › Hard coding and STEM
  • › Deep research

Strengths

  • › State-of-the-art agentic workflows
  • › Complex STEM reasoning
  • › Sol Ultra mode for highest effort

What usually breaks

  • ! Government-gated preview access limitations
  • ! High cost for Ultra mode reasoning
  • ! Still needs grounding for extremely long contexts
Docs / source →

GPT-5.6 Terra

OpenAI ·

APIMultimodalReasoning
Context Window
Large
License
Proprietary (API)
Pricing
$2 input / $12 output per 1M
Fine-tune
noN/A

Best for

  • › Everyday coding
  • › Data extraction
  • › Routing

Strengths

  • › Balanced everyday workhorse
  • › Strong cost-to-performance ratio
  • › Fast multi-agent coordination

What usually breaks

  • ! Not the absolute SOTA for extreme STEM
  • ! Preview restrictions apply
Docs / source →

DeepSeek V4

DeepSeek ·

Open weightsReasoning
Context Window
128K
License
Open weights
Pricing
Free weights; API peak $1.32 / $3.96 per 1M (Pro)
Fine-tune
yesStandard tooling supports fine-tuning well.

Best for

  • › Cost-sensitive reasoning
  • › Coding benchmarks

Strengths

  • › Value king for reasoning
  • › 1.6T parameters
  • › Aggressive pricing

What usually breaks

  • ! Overtaken by Kimi K3 for top open-weight capability
  • ! License specifics for commercial use
Docs / source →

Gemini 3.1 Pro

Google ·

APIMultimodalReasoning
Context Window
1M
License
Proprietary (API)
Pricing
$2 input / $12 output per 1M (<=200K prompts)
Fine-tune
limitedVertex AI fine-tuning.

Best for

  • › Long-document analysis
  • › Deep research
  • › GCP native apps

Strengths

  • › Current public Gemini Pro SKU
  • › Strong multimodal + agentic work
  • › Google ecosystem integration

What usually breaks

  • ! Still a preview ID (gemini-3.1-pro-preview)
  • ! Prompts >200K jump to $4 / $18
  • ! Gemini 3.5 Pro is announced but not publicly priced
Docs / source →

Llama 3.3 70B

Meta ·

Open weights
Context Window
128K
License
Llama Community License
Pricing
Free weights
Fine-tune
yesFull fine-tune / LoRA common.

Best for

  • › Private deployments
  • › Cost control

Strengths

  • › Reliable mid-tier open model
  • › VPC friendly

What usually breaks

  • ! No deep reasoning loop
  • ! Replaced by newer massive models (K3, Inkling) at the high end
Docs / source →

GPT-4o

OpenAI ·

APIMultimodal
Context Window
128K
License
Proprietary (API)
Pricing
Legacy pricing
Fine-tune
limitedFine-tuning available on select GPT-4o variants.

Best for

  • › Fallback systems
  • › Legacy workflows

Strengths

  • › General chat & coding
  • › Vision + audio product surface

What usually breaks

  • ! Legacy model; superseded by 5.5 and 5.6
Docs / source →